{
  "benchmark_line": "gamebench2",
  "comparability_tuples": [
    {
      "evaluation_family": "codegen_public_benchmark",
      "rating_method": "sequential_elo",
      "reasoning_effort": "highest",
      "results_cut_id": "20260720_093741",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_kind": "full_freedom"
    },
    {
      "evaluation_family": "codegen_public_benchmark",
      "rating_method": "sequential_elo",
      "reasoning_effort": "medium",
      "results_cut_id": "20260720_093741",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_kind": "full_freedom"
    },
    {
      "evaluation_family": "codegen_public_benchmark",
      "rating_method": "sequential_elo",
      "reasoning_effort": "mixed",
      "results_cut_id": "20260720_093741",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_kind": "full_freedom"
    },
    {
      "evaluation_family": "codegen_public_benchmark",
      "rating_method": "sequential_elo",
      "reasoning_effort": "none",
      "results_cut_id": "20260720_093741",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_kind": "full_freedom"
    },
    {
      "evaluation_family": "codegen_public_benchmark",
      "rating_method": "sequential_elo",
      "reasoning_effort": "aggregate",
      "results_cut_id": "20260720_093741",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_kind": "overall"
    }
  ],
  "generated_at": "2026-07-20T09:37:41Z",
  "public_release_id": "gb2-20260720-959ae35a724b",
  "results_cut_id": "gb2-20260720-959ae35a724b",
  "rows": [
    {
      "admitted_plugins": 8,
      "attempted_matches": 1379,
      "capped_draws": 1,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 11
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 1.91725,
        "mean": 1.58262125,
        "median": 1.5854,
        "min": 1.05322,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 11
        },
        "run_count": 8,
        "token_priced_attempt_count": 11,
        "total": 12.66097
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 89.02939366220986,
      "contributing_entrants": null,
      "distinct_rated_opponents": 652,
      "draws": 129,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04"
      ],
      "infra_failures_excluded": 4,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:36Z",
      "losses": 98,
      "max_score": 98.290012518652,
      "measured_at": null,
      "min_score": 72.93980546270726,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": 1.0,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1343,
      "rating_method": "sequential_elo",
      "raw_elo": 1897.6533687768751,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 89.02939366220986,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 35,
      "wins": 1116
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1469,
      "capped_draws": 29,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 12.915325,
        "mean": 10.059243749999998,
        "median": 10.0094025,
        "min": 8.401555,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 80.47395
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 70.61851949173621,
      "contributing_entrants": null,
      "distinct_rated_opponents": 656,
      "draws": 135,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 8,
      "first_pass_rate": 1.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:55Z",
      "losses": 288,
      "max_score": 100.0,
      "measured_at": null,
      "min_score": 7.602053478392984,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1439,
      "rating_method": "sequential_elo",
      "raw_elo": 1778.6799910644756,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 70.61851949173621,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 15,
      "wins": 1016
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1355,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.6981,
        "mean": 0.5046499999999999,
        "median": 0.511355,
        "min": 0.33027,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 4.0372
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 71.77619468353247,
      "contributing_entrants": null,
      "distinct_rated_opponents": 628,
      "draws": 181,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:02Z",
      "losses": 137,
      "max_score": 85.85876169674393,
      "measured_at": null,
      "min_score": 44.00470514710234,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": 0.875,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1349,
      "rating_method": "sequential_elo",
      "raw_elo": 1862.005018618128,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 69.41965374228975,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 1031
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1314,
      "capped_draws": 111,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 4.4173,
        "mean": 2.7356843750000004,
        "median": 2.536595,
        "min": 2.045485,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 21.885474999999996
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 69.42492057631782,
      "contributing_entrants": null,
      "distinct_rated_opponents": 590,
      "draws": 264,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-02"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:21Z",
      "losses": 154,
      "max_score": 100.0,
      "measured_at": null,
      "min_score": 19.24432865630252,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": 0.875,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1287,
      "rating_method": "sequential_elo",
      "raw_elo": 1834.8406889372725,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 67.14557617248093,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 12,
      "wins": 869
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1488,
      "capped_draws": 44,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 3.2933475,
        "mean": 2.642580625,
        "median": 2.681945,
        "min": 1.562975,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 21.140645
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 64.66453134183178,
      "contributing_entrants": null,
      "distinct_rated_opponents": 666,
      "draws": 263,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:05Z",
      "losses": 122,
      "max_score": 100.0,
      "measured_at": null,
      "min_score": 9.980876737333936,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1454,
      "rating_method": "sequential_elo",
      "raw_elo": 1860.100711513732,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 64.66453134183178,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 15,
      "wins": 1069
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1462,
      "capped_draws": 18,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 10
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.6689649999999999,
        "mean": 0.46799875,
        "median": 0.4352825,
        "min": 0.366615,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 10
        },
        "run_count": 9,
        "token_priced_attempt_count": 10,
        "total": 4.1199200000000005
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 64.64087807182851,
      "contributing_entrants": null,
      "distinct_rated_opponents": 639,
      "draws": 247,
      "entry_count": 8,
      "failed_game_aliases": [
        "GB2-06"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.7777777777777778,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:39:18Z",
      "losses": 177,
      "max_score": 90.58621353721448,
      "measured_at": null,
      "min_score": 25.456448781739248,
      "model_code_failure_rate": 0.1111111111111111,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": 0.8888888888888888,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1259,
      "rating_method": "sequential_elo",
      "raw_elo": 1764.368616883458,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.1111111111111111,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 62.76522883961819,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 195,
      "wins": 835
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1395,
      "capped_draws": 70,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 11
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 12.13752,
        "mean": 8.1660015625,
        "median": 8.18098,
        "min": 4.7142525,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 11
        },
        "run_count": 8,
        "token_priced_attempt_count": 11,
        "total": 65.32801249999999
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 57.60709859351452,
      "contributing_entrants": null,
      "distinct_rated_opponents": 656,
      "draws": 301,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:22Z",
      "losses": 272,
      "max_score": 100.0,
      "measured_at": null,
      "min_score": 23.297513470769697,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1373,
      "rating_method": "sequential_elo",
      "raw_elo": 1673.782357430351,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 57.60709859351452,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 16,
      "wins": 800
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1307,
      "capped_draws": 11,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 10
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 4.987857,
        "mean": 3.1645696249999995,
        "median": 3.0710854999999997,
        "min": 1.985613,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 10
        },
        "run_count": 9,
        "token_priced_attempt_count": 10,
        "total": 28.806541000000003
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 54.66589143146764,
      "contributing_entrants": null,
      "distinct_rated_opponents": 653,
      "draws": 298,
      "entry_count": 8,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.7777777777777778,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:07Z",
      "losses": 211,
      "max_score": 85.25976029677969,
      "measured_at": null,
      "min_score": 9.848517519974454,
      "model_code_failure_rate": 0.1111111111111111,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": 0.8888888888888888,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1291,
      "rating_method": "sequential_elo",
      "raw_elo": 1702.2029115290434,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.1111111111111111,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 53.07968096604684,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 10,
      "wins": 782
    },
    {
      "admitted_plugins": 14,
      "attempted_matches": 899,
      "capped_draws": 92,
      "codegen_cost_usd": {
        "admitted_cell_count": 14,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 34
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 1.1371809,
        "mean": 0.8062893374999999,
        "median": 0.81347715,
        "min": 0.5199072,
        "recorded_attempt_count": 34,
        "run_count": 16,
        "total": 12.900629399999998
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": 54.22864747461619,
      "contributing_entrants": null,
      "distinct_rated_opponents": 418,
      "draws": 228,
      "entry_count": 14,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-02",
        "GB2-08"
      ],
      "infra_failures_excluded": 2,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:15Z",
      "losses": 267,
      "max_score": 82.92666095797935,
      "measured_at": null,
      "min_score": 23.77593276686914,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 2,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": 0.875,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 877,
      "rating_method": "sequential_elo",
      "raw_elo": 1610.2509525723651,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 14,
      "repaired_rate": 0.875,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 52.44822391600324,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 22,
      "wins": 382
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1414,
      "capped_draws": 126,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 19
        },
        "evidence_basis": "Provider-reported",
        "max": 0.5145615,
        "mean": 0.34875300000000004,
        "median": 0.335163,
        "min": 0.254883,
        "recorded_attempt_count": 19,
        "run_count": 8,
        "total": 2.790024
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 52.16906430603814,
      "contributing_entrants": null,
      "distinct_rated_opponents": 658,
      "draws": 415,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:33Z",
      "losses": 296,
      "max_score": 91.79025106882118,
      "measured_at": null,
      "min_score": 15.64390388899189,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": 1.0,
      "provider": "Google",
      "provider_key": "google",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1390,
      "rating_method": "sequential_elo",
      "raw_elo": 1667.2044190421009,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 8,
      "repaired_rate": 1.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 52.16906430603814,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 13,
      "wins": 679
    },
    {
      "admitted_plugins": 15,
      "attempted_matches": 834,
      "capped_draws": 20,
      "codegen_cost_usd": {
        "admitted_cell_count": 15,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 30
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.141566,
        "mean": 0.10599,
        "median": 0.105677,
        "min": 0.046481,
        "recorded_attempt_count": 30,
        "run_count": 16,
        "total": 1.69584
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": 51.00775136236891,
      "contributing_entrants": null,
      "distinct_rated_opponents": 302,
      "draws": 141,
      "entry_count": 15,
      "failed_game_aliases": [
        "GB2-02"
      ],
      "first_pass_admitted": 3,
      "first_pass_rate": 0.1875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:42Z",
      "losses": 95,
      "max_score": 66.83613208021495,
      "measured_at": null,
      "min_score": 15.80923544929635,
      "model_code_failure_rate": 0.0625,
      "model_code_failures": 1,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": 0.9375,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 604,
      "rating_method": "sequential_elo",
      "raw_elo": 1664.1613430404243,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 12,
      "repaired_rate": 0.75,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 50.191363917808594,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 222,
      "wins": 368
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1466,
      "capped_draws": 22,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 1.38945,
        "mean": 0.93123625,
        "median": 0.95991,
        "min": 0.36561,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 7.449890000000001
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 51.47924961758838,
      "contributing_entrants": null,
      "distinct_rated_opponents": 650,
      "draws": 172,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-04",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 6,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:43:48Z",
      "losses": 230,
      "max_score": 51.47924961758838,
      "measured_at": null,
      "min_score": 51.47924961758838,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": 0.875,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1458,
      "rating_method": "sequential_elo",
      "raw_elo": 1828.7659874140832,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 49.78909371167585,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 8,
      "wins": 1056
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1599,
      "capped_draws": 127,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 10
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.1598575,
        "mean": 0.093565,
        "median": 0.07461,
        "min": 0.06826499999999999,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 10
        },
        "run_count": 8,
        "token_priced_attempt_count": 10,
        "total": 0.7485199999999999
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 48.68896152233137,
      "contributing_entrants": null,
      "distinct_rated_opponents": 657,
      "draws": 446,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:24Z",
      "losses": 293,
      "max_score": 91.35944335010576,
      "measured_at": null,
      "min_score": 11.529384168381274,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1385,
      "rating_method": "sequential_elo",
      "raw_elo": 1642.005780104476,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.25,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 48.68896152233137,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 204,
      "wins": 646
    },
    {
      "admitted_plugins": 12,
      "attempted_matches": 2345,
      "capped_draws": 169,
      "codegen_cost_usd": {
        "admitted_cell_count": 12,
        "cell_count": 7,
        "cost_basis_counts": {
          "local_standard_estimate": 15
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 1.5164726249999998,
        "mean": 1.0182428035714286,
        "median": 0.9670455,
        "min": 0.8356068750000001,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 15
        },
        "run_count": 12,
        "token_priced_attempt_count": 15,
        "total": 12.214044
      },
      "codegen_evaluated_cells": 12,
      "conditional_success_score": 48.540299494544975,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1097,
      "draws": 616,
      "entry_count": 12,
      "failed_game_aliases": [],
      "first_pass_admitted": 10,
      "first_pass_rate": 0.8333333333333334,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-04"
      ],
      "infra_failures_excluded": 6,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:47Z",
      "losses": 606,
      "max_score": 80.51515814839607,
      "measured_at": null,
      "min_score": 12.309339079481154,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 2332,
      "rating_method": "sequential_elo",
      "raw_elo": 1614.6439683123333,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.16666666666666666,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 48.540299494544975,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 13,
      "wins": 1110
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1455,
      "capped_draws": 69,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.712055,
        "mean": 0.49864374999999994,
        "median": 0.5065824999999999,
        "min": 0.29423750000000004,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 3.98915
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 48.01757453308687,
      "contributing_entrants": null,
      "distinct_rated_opponents": 625,
      "draws": 388,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:45Z",
      "losses": 261,
      "max_score": 99.15475058347155,
      "measured_at": null,
      "min_score": 14.741184593734893,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1277,
      "rating_method": "sequential_elo",
      "raw_elo": 1644.2583933799935,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 48.01757453308687,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 178,
      "wins": 628
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1407,
      "capped_draws": 130,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.144235,
        "mean": 0.12174125,
        "median": 0.121475,
        "min": 0.09594,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 0.97393
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 45.67057735948555,
      "contributing_entrants": null,
      "distinct_rated_opponents": 660,
      "draws": 413,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 8,
      "first_pass_rate": 1.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:59Z",
      "losses": 388,
      "max_score": 78.90896540288884,
      "measured_at": null,
      "min_score": 11.482980464184529,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": 1.0,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1391,
      "rating_method": "sequential_elo",
      "raw_elo": 1573.4311153668432,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 45.67057735948555,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 7,
      "wins": 590
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1561,
      "capped_draws": 232,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.58673,
        "mean": 0.442856875,
        "median": 0.4176025,
        "min": 0.34269,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 3.5428550000000003
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 45.05538252246229,
      "contributing_entrants": null,
      "distinct_rated_opponents": 652,
      "draws": 481,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:44:30Z",
      "losses": 268,
      "max_score": 81.04306968418157,
      "measured_at": null,
      "min_score": 20.085766120064036,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1359,
      "rating_method": "sequential_elo",
      "raw_elo": 1584.536956534744,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 45.05538252246229,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 199,
      "wins": 610
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1309,
      "capped_draws": 142,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 10
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.454645,
        "mean": 0.28201406249999994,
        "median": 0.2695775,
        "min": 0.17401125,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 10
        },
        "run_count": 8,
        "token_priced_attempt_count": 10,
        "total": 2.2561125
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 44.65283798152629,
      "contributing_entrants": null,
      "distinct_rated_opponents": 656,
      "draws": 429,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:07Z",
      "losses": 305,
      "max_score": 82.16527921865402,
      "measured_at": null,
      "min_score": 14.429040501509393,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1293,
      "rating_method": "sequential_elo",
      "raw_elo": 1591.4589626214092,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.25,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 44.65283798152629,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 10,
      "wins": 559
    },
    {
      "admitted_plugins": 15,
      "attempted_matches": 2806,
      "capped_draws": 389,
      "codegen_cost_usd": {
        "admitted_cell_count": 15,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 25
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 2,
        "max": 0.38568888333333334,
        "mean": 0.29249454166666666,
        "median": 0.2834616125,
        "min": 0.2238526,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 25
        },
        "run_count": 17,
        "token_priced_attempt_count": 25,
        "total": 5.07700365
      },
      "codegen_evaluated_cells": 17,
      "conditional_success_score": 45.64596858389405,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1292,
      "draws": 852,
      "entry_count": 15,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 10,
      "first_pass_rate": 0.5882352941176471,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07"
      ],
      "infra_failures_excluded": 9,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:39:21Z",
      "losses": 851,
      "max_score": 45.64596858389405,
      "measured_at": null,
      "min_score": 45.64596858389405,
      "model_code_failure_rate": 0.1176470588235294,
      "model_code_failures": 2,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": 0.8823529411764706,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 2775,
      "rating_method": "sequential_elo",
      "raw_elo": 1588.1684458982304,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.29411764705882354,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 44.23978538225884,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 22,
      "wins": 1072
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1403,
      "capped_draws": 170,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.34174499999999997,
        "mean": 0.217940625,
        "median": 0.19638,
        "min": 0.14813,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 1.743525
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 44.07615889983954,
      "contributing_entrants": null,
      "distinct_rated_opponents": 657,
      "draws": 491,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 8,
      "first_pass_rate": 1.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:39:26Z",
      "losses": 322,
      "max_score": 91.0317105229074,
      "measured_at": null,
      "min_score": 11.892584996054628,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": 1.0,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1387,
      "rating_method": "sequential_elo",
      "raw_elo": 1606.1167124424544,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 44.07615889983954,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 8,
      "wins": 574
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1366,
      "capped_draws": 41,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.190125,
        "mean": 0.15785312499999998,
        "median": 0.15388000000000002,
        "min": 0.131495,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 1.262825
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 44.86404683281706,
      "contributing_entrants": null,
      "distinct_rated_opponents": 638,
      "draws": 335,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:02Z",
      "losses": 435,
      "max_score": 80.76792349183744,
      "measured_at": null,
      "min_score": 13.717353446889966,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": 0.875,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1366,
      "rating_method": "sequential_elo",
      "raw_elo": 1583.4197099575747,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 43.391079874656214,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 596
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 967,
      "capped_draws": 106,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 18
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.6921525,
        "mean": 0.4890770625000001,
        "median": 0.483741,
        "min": 0.401826,
        "recorded_attempt_count": 18,
        "run_count": 8,
        "total": 3.912616499999999
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 45.710505493521154,
      "contributing_entrants": null,
      "distinct_rated_opponents": 475,
      "draws": 188,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:20Z",
      "losses": 297,
      "max_score": 87.65213640227793,
      "measured_at": null,
      "min_score": 5.549046286100777,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": 0.75,
      "provider": "Google",
      "provider_key": "google",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 932,
      "rating_method": "sequential_elo",
      "raw_elo": 1587.3523598129802,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.75,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 42.538418524284005,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 33,
      "wins": 447
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1260,
      "capped_draws": 67,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.3825,
        "mean": 0.27230375,
        "median": 0.25636749999999997,
        "min": 0.236235,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 2.17843
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 43.09225868476076,
      "contributing_entrants": null,
      "distinct_rated_opponents": 630,
      "draws": 350,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:22Z",
      "losses": 376,
      "max_score": 71.23639433663045,
      "measured_at": null,
      "min_score": 13.07288327052516,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": 0.875,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1259,
      "rating_method": "sequential_elo",
      "raw_elo": 1574.9480374312168,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 41.67746270276425,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 533
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1079,
      "capped_draws": 156,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 16
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.10909930499999998,
        "mean": 0.082847753875,
        "median": 0.084146916,
        "min": 0.03951746,
        "recorded_attempt_count": 16,
        "run_count": 8,
        "total": 0.662782031
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 44.70871398411597,
      "contributing_entrants": null,
      "distinct_rated_opponents": 539,
      "draws": 265,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 1,
      "first_pass_rate": 0.125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:09Z",
      "losses": 343,
      "max_score": 85.75114955106763,
      "measured_at": null,
      "min_score": 14.985507450020902,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": 0.75,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1078,
      "rating_method": "sequential_elo",
      "raw_elo": 1567.0799686218386,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 41.60614647782431,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 470
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1211,
      "capped_draws": 144,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 2,
        "max": 0.072978,
        "mean": 0.048872,
        "median": 0.046325000000000005,
        "min": 0.03731,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 0.390976
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 44.64232956611945,
      "contributing_entrants": null,
      "distinct_rated_opponents": 571,
      "draws": 291,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:00Z",
      "losses": 402,
      "max_score": 64.12287600481088,
      "measured_at": null,
      "min_score": 21.75260522744053,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": 0.75,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1209,
      "rating_method": "sequential_elo",
      "raw_elo": 1558.5916787756134,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 41.54436881586808,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 1,
      "wins": 516
    },
    {
      "admitted_plugins": 13,
      "attempted_matches": 524,
      "capped_draws": 12,
      "codegen_cost_usd": {
        "admitted_cell_count": 13,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 26
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.184451,
        "mean": 0.0949985,
        "median": 0.08180399999999999,
        "min": 0.050682000000000005,
        "recorded_attempt_count": 26,
        "run_count": 16,
        "total": 1.5199759999999998
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": 43.36499484850174,
      "contributing_entrants": null,
      "distinct_rated_opponents": 262,
      "draws": 104,
      "entry_count": 13,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.4375,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:02Z",
      "losses": 139,
      "max_score": 59.14636993260279,
      "measured_at": null,
      "min_score": 11.001592922177778,
      "model_code_failure_rate": 0.1875,
      "model_code_failures": 3,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": 0.8125,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 524,
      "rating_method": "sequential_elo",
      "raw_elo": 1604.4421371991327,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 41.17135321287196,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 281
    },
    {
      "admitted_plugins": 10,
      "attempted_matches": 1587,
      "capped_draws": 154,
      "codegen_cost_usd": {
        "admitted_cell_count": 10,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 10
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.4567,
        "mean": 0.3660333333333334,
        "median": 0.355077,
        "min": 0.234622,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 10
        },
        "run_count": 10,
        "token_priced_attempt_count": 10,
        "total": 3.5443080000000005
      },
      "codegen_evaluated_cells": 10,
      "conditional_success_score": 40.57860418225832,
      "contributing_entrants": null,
      "distinct_rated_opponents": 673,
      "draws": 444,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 10,
      "first_pass_rate": 1.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 18,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:21Z",
      "losses": 435,
      "max_score": 64.79650419125173,
      "measured_at": null,
      "min_score": 0.0,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": 1.0,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1524,
      "rating_method": "sequential_elo",
      "raw_elo": 1568.0429525721834,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 40.57860418225832,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 53,
      "wins": 645
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1203,
      "capped_draws": 25,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 18
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.226689264,
        "mean": 0.16037539425,
        "median": 0.140728509,
        "min": 0.10689655,
        "recorded_attempt_count": 18,
        "run_count": 8,
        "total": 1.283003154
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 43.5273049102168,
      "contributing_entrants": null,
      "distinct_rated_opponents": 554,
      "draws": 309,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-06"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T14:35:59Z",
      "losses": 361,
      "max_score": 59.41026994182745,
      "measured_at": null,
      "min_score": 17.581567025013598,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": 0.75,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1201,
      "rating_method": "sequential_elo",
      "raw_elo": 1596.5699667703082,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.75,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 40.50672145306643,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 531
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1453,
      "capped_draws": 146,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.046574,
        "mean": 0.036155625,
        "median": 0.035338,
        "min": 0.025803,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 9,
        "token_priced_attempt_count": 9,
        "total": 0.315048
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 40.92861802674063,
      "contributing_entrants": null,
      "distinct_rated_opponents": 645,
      "draws": 442,
      "entry_count": 8,
      "failed_game_aliases": [
        "GB2-06"
      ],
      "first_pass_admitted": 8,
      "first_pass_rate": 0.8888888888888888,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:14Z",
      "losses": 345,
      "max_score": 73.91132063132736,
      "measured_at": null,
      "min_score": 11.282582150463083,
      "model_code_failure_rate": 0.1111111111111111,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": 0.8888888888888888,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1273,
      "rating_method": "sequential_elo",
      "raw_elo": 1549.3497212912257,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 39.74101455866921,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 177,
      "wins": 486
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 965,
      "capped_draws": 128,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 17
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.2996388,
        "mean": 0.13471628689999998,
        "median": 0.1144847257,
        "min": 0.0276132876,
        "recorded_attempt_count": 17,
        "run_count": 8,
        "total": 1.0777302952
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 42.4596419290968,
      "contributing_entrants": null,
      "distinct_rated_opponents": 460,
      "draws": 247,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:39:33Z",
      "losses": 429,
      "max_score": 68.412551151382,
      "measured_at": null,
      "min_score": 13.36353157637584,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": 0.75,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 30,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 965,
      "rating_method": "sequential_elo",
      "raw_elo": 1523.4278723519426,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.75,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 39.51314909495272,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 492
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1492,
      "capped_draws": 79,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.147405,
        "mean": 0.121464375,
        "median": 0.124865,
        "min": 0.09045,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 9,
        "token_priced_attempt_count": 9,
        "total": 1.0694850000000002
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 40.136337833904406,
      "contributing_entrants": null,
      "distinct_rated_opponents": 653,
      "draws": 348,
      "entry_count": 8,
      "failed_game_aliases": [
        "GB2-04"
      ],
      "first_pass_admitted": 8,
      "first_pass_rate": 0.8888888888888888,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:29Z",
      "losses": 405,
      "max_score": 61.398138957666426,
      "measured_at": null,
      "min_score": 14.58766401287794,
      "model_code_failure_rate": 0.1111111111111111,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": 0.8888888888888888,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 31,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1288,
      "rating_method": "sequential_elo",
      "raw_elo": 1528.1411119357276,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 38.971723529651854,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 196,
      "wins": 535
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1329,
      "capped_draws": 188,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.309345,
        "mean": 0.239705,
        "median": 0.232815,
        "min": 0.175235,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 1.91764
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 38.86135552659715,
      "contributing_entrants": null,
      "distinct_rated_opponents": 629,
      "draws": 429,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 8,
      "first_pass_rate": 1.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:38Z",
      "losses": 399,
      "max_score": 61.54507370770128,
      "measured_at": null,
      "min_score": 14.385271315724854,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": 1.0,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 32,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1304,
      "rating_method": "sequential_elo",
      "raw_elo": 1527.9629146538243,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 38.86135552659715,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 16,
      "wins": 476
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1198,
      "capped_draws": 181,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 16
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.08827968,
        "mean": 0.05437612,
        "median": 0.0574344,
        "min": 0.00733824,
        "recorded_attempt_count": 16,
        "run_count": 8,
        "total": 0.43500896
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 41.3850467207755,
      "contributing_entrants": null,
      "distinct_rated_opponents": 552,
      "draws": 367,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-04"
      ],
      "first_pass_admitted": 1,
      "first_pass_rate": 0.125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T15:12:19Z",
      "losses": 378,
      "max_score": 78.90361697494383,
      "measured_at": null,
      "min_score": 12.24169582676053,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": 0.75,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 33,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1198,
      "rating_method": "sequential_elo",
      "raw_elo": 1531.879104781743,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 38.51312557252109,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 453
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1226,
      "capped_draws": 147,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.21477375,
        "mean": 0.17756265624999998,
        "median": 0.17285124999999998,
        "min": 0.1418975,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 1.4205012499999998
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 39.65661326048428,
      "contributing_entrants": null,
      "distinct_rated_opponents": 585,
      "draws": 279,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-02"
      ],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:37Z",
      "losses": 465,
      "max_score": 68.30589565602956,
      "measured_at": null,
      "min_score": 15.994049148715748,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": 0.875,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 34,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1223,
      "rating_method": "sequential_elo",
      "raw_elo": 1504.5443847461988,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 38.35461566711227,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 479
    },
    {
      "admitted_plugins": 16,
      "attempted_matches": 2576,
      "capped_draws": 394,
      "codegen_cost_usd": {
        "admitted_cell_count": 16,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 46
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 8,
        "max": 0.16378006679999998,
        "mean": 0.12038942953333334,
        "median": 0.1261480173333333,
        "min": 0.05546312440000001,
        "recorded_attempt_count": 46,
        "run_count": 24,
        "total": 2.8893463088
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": 41.131076513149246,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1213,
      "draws": 723,
      "entry_count": 16,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06"
      ],
      "first_pass_admitted": 3,
      "first_pass_rate": 0.125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:57Z",
      "losses": 624,
      "max_score": 64.67108012218564,
      "measured_at": null,
      "min_score": 17.591072904112846,
      "model_code_failure_rate": 0.3333333333333333,
      "model_code_failures": 8,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": 0.6666666666666666,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 35,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 2342,
      "rating_method": "sequential_elo",
      "raw_elo": 1568.120841782094,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 13,
      "repaired_rate": 0.5416666666666666,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 37.166123147911485,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 212,
      "wins": 995
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1054,
      "capped_draws": 61,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 13
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.09803328,
        "mean": 0.0253809975,
        "median": 0.00943521,
        "min": 0.00639384,
        "recorded_attempt_count": 13,
        "run_count": 8,
        "total": 0.20304798
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 39.60055944956829,
      "contributing_entrants": null,
      "distinct_rated_opponents": 486,
      "draws": 238,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-04",
        "GB2-06"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.5,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:59:28Z",
      "losses": 412,
      "max_score": 54.04543152948865,
      "measured_at": null,
      "min_score": 6.05124929753971,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": 0.75,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 36,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1041,
      "rating_method": "sequential_elo",
      "raw_elo": 1517.3290306459578,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.25,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 36.852473046929816,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 4,
      "wins": 391
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1129,
      "capped_draws": 179,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 9
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.03912625,
        "mean": 0.022799843750000003,
        "median": 0.0216425,
        "min": 0.015285,
        "recorded_attempt_count": 9,
        "run_count": 8,
        "total": 0.18239875
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 37.35756527528258,
      "contributing_entrants": null,
      "distinct_rated_opponents": 538,
      "draws": 275,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-04"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T12:57:38Z",
      "losses": 379,
      "max_score": 70.14174273877752,
      "measured_at": null,
      "min_score": 11.919704676069871,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": 0.875,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 37,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1109,
      "rating_method": "sequential_elo",
      "raw_elo": 1500.1511607756422,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 36.13104954225295,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 10,
      "wins": 455
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1510,
      "capped_draws": 158,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 11
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.24673750000000003,
        "mean": 0.16744,
        "median": 0.158095,
        "min": 0.108735,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 11
        },
        "run_count": 8,
        "token_priced_attempt_count": 11,
        "total": 1.33952
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 35.910461569583475,
      "contributing_entrants": null,
      "distinct_rated_opponents": 660,
      "draws": 471,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:31Z",
      "losses": 373,
      "max_score": 66.13437229960988,
      "measured_at": null,
      "min_score": 10.838789505864996,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 38,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1306,
      "rating_method": "sequential_elo",
      "raw_elo": 1506.2791701522935,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.910461569583475,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 195,
      "wins": 462
    },
    {
      "admitted_plugins": 19,
      "attempted_matches": 3362,
      "capped_draws": 418,
      "codegen_cost_usd": {
        "admitted_cell_count": 19,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 57
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 8,
        "max": 0.25293804045,
        "mean": 0.19953128278325,
        "median": 0.20662718916666667,
        "min": 0.13279406625,
        "recorded_attempt_count": 57,
        "run_count": 27,
        "total": 5.24138213703
      },
      "codegen_evaluated_cells": 27,
      "conditional_success_score": 38.86944403772714,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1586,
      "draws": 1101,
      "entry_count": 18,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-04",
        "GB2-06"
      ],
      "first_pass_admitted": 1,
      "first_pass_rate": 0.037037037037037035,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-03",
        "GB2-06"
      ],
      "infra_failures_excluded": 2,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:49:57Z",
      "losses": 929,
      "max_score": 62.93805242641539,
      "measured_at": null,
      "min_score": 9.910010721360797,
      "model_code_failure_rate": 0.2962962962962963,
      "model_code_failures": 8,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": 0.7037037037037037,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 39,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 3287,
      "rating_method": "sequential_elo",
      "raw_elo": 1557.8020900682966,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 18,
      "repaired_rate": 0.6666666666666666,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.600474625013355,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 49,
      "wins": 1257
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1205,
      "capped_draws": 218,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 12
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.029078938,
        "mean": 0.015562441375,
        "median": 0.0131167185,
        "min": 0.005612805,
        "recorded_attempt_count": 12,
        "run_count": 8,
        "total": 0.124499531
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 36.77193379881386,
      "contributing_entrants": null,
      "distinct_rated_opponents": 562,
      "draws": 404,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-04"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.5,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:22:40Z",
      "losses": 396,
      "max_score": 75.05378217838995,
      "measured_at": null,
      "min_score": 10.465061755242804,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": 0.875,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 40,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1189,
      "rating_method": "sequential_elo",
      "raw_elo": 1491.2315536146675,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.56464539535866,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 7,
      "wins": 389
    },
    {
      "admitted_plugins": 28,
      "attempted_matches": 4908,
      "capped_draws": 240,
      "codegen_cost_usd": {
        "admitted_cell_count": 28,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 1,
          "provider_reported": 47
        },
        "evidence_basis": "Mixed: Provider-reported + Standard list-price estimate",
        "failed_cell_count": 4,
        "max": 0.0370864994,
        "mean": 0.02414278511875,
        "median": 0.021930847225,
        "min": 0.0140238704,
        "recorded_attempt_count": 47,
        "run_count": 32,
        "token_priced_attempt_count": 1,
        "total": 0.7725691238000001
      },
      "codegen_evaluated_cells": 32,
      "conditional_success_score": 36.77177511539484,
      "contributing_entrants": null,
      "distinct_rated_opponents": 2263,
      "draws": 1243,
      "entry_count": 28,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-07",
        "GB2-08"
      ],
      "first_pass_admitted": 16,
      "first_pass_rate": 0.5,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:44:44Z",
      "losses": 1634,
      "max_score": 47.55645299565777,
      "measured_at": null,
      "min_score": 23.89158662732705,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 4,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": 0.875,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 41,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 4661,
      "rating_method": "sequential_elo",
      "raw_elo": 1510.502330589102,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 12,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.56449192180031,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 218,
      "wins": 1784
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1368,
      "capped_draws": 170,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.132112,
        "mean": 0.0780535,
        "median": 0.06797,
        "min": 0.05673,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 0.624428
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 36.75508899972182,
      "contributing_entrants": null,
      "distinct_rated_opponents": 637,
      "draws": 458,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:07Z",
      "losses": 419,
      "max_score": 66.50406286570629,
      "measured_at": null,
      "min_score": 10.164730563488815,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": 0.875,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 42,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1364,
      "rating_method": "sequential_elo",
      "raw_elo": 1509.2425222105776,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.54835364117075,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 487
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1378,
      "capped_draws": 179,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 15
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.121855,
        "mean": 0.09110296875,
        "median": 0.09280125,
        "min": 0.0343775,
        "recorded_attempt_count": 15,
        "run_count": 8,
        "total": 0.7288237499999999
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 36.66582431412571,
      "contributing_entrants": null,
      "distinct_rated_opponents": 642,
      "draws": 464,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 1,
      "first_pass_rate": 0.125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T14:28:19Z",
      "losses": 435,
      "max_score": 76.51799749682782,
      "measured_at": null,
      "min_score": 12.808440532491185,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": 0.875,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 43,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1377,
      "rating_method": "sequential_elo",
      "raw_elo": 1514.0501560831765,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.75,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.4620196749746,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 478
    },
    {
      "admitted_plugins": 17,
      "attempted_matches": 3169,
      "capped_draws": 359,
      "codegen_cost_usd": {
        "admitted_cell_count": 17,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 35
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 7,
        "max": 0.004834030033333333,
        "mean": 0.0027724950333333333,
        "median": 0.0022439992166666664,
        "min": 0.00169122,
        "recorded_attempt_count": 35,
        "run_count": 24,
        "total": 0.06653988079999999
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": 38.57796942583274,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1471,
      "draws": 782,
      "entry_count": 17,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-07",
        "GB2-08"
      ],
      "first_pass_admitted": 12,
      "first_pass_rate": 0.5,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:38:40Z",
      "losses": 1468,
      "max_score": 39.07689302053589,
      "measured_at": null,
      "min_score": 37.704641705021,
      "model_code_failure_rate": 0.2916666666666667,
      "model_code_failures": 7,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": 0.7083333333333334,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 44,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 3131,
      "rating_method": "sequential_elo",
      "raw_elo": 1387.3705081096236,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.20833333333333337,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.39148490151478,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 35,
      "wins": 881
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1412,
      "capped_draws": 167,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.07806,
        "mean": 0.06121625,
        "median": 0.061152,
        "min": 0.049788,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 0.48973
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 34.30827338646576,
      "contributing_entrants": null,
      "distinct_rated_opponents": 659,
      "draws": 487,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 8,
      "first_pass_rate": 1.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01"
      ],
      "infra_failures_excluded": 1,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:44:37Z",
      "losses": 404,
      "max_score": 60.31532930836284,
      "measured_at": null,
      "min_score": 11.593620745823015,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": 1.0,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 45,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1392,
      "rating_method": "sequential_elo",
      "raw_elo": 1520.2481939112397,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 34.30827338646576,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 5,
      "wins": 501
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1407,
      "capped_draws": 134,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 11
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.094594,
        "mean": 0.0571955,
        "median": 0.043012,
        "min": 0.032534,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 11
        },
        "run_count": 8,
        "token_priced_attempt_count": 11,
        "total": 0.457564
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 34.16293015239342,
      "contributing_entrants": null,
      "distinct_rated_opponents": 664,
      "draws": 445,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:41Z",
      "losses": 511,
      "max_score": 57.80317048612072,
      "measured_at": null,
      "min_score": 8.667504440488187,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 46,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1396,
      "rating_method": "sequential_elo",
      "raw_elo": 1474.9462497663255,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 34.16293015239342,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 3,
      "wins": 440
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1489,
      "capped_draws": 188,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 10
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.1192425,
        "mean": 0.0683559375,
        "median": 0.06152125,
        "min": 0.0542775,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 10
        },
        "run_count": 9,
        "token_priced_attempt_count": 10,
        "total": 0.6063375
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 34.051070978069944,
      "contributing_entrants": null,
      "distinct_rated_opponents": 651,
      "draws": 499,
      "entry_count": 8,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.7777777777777778,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:09Z",
      "losses": 434,
      "max_score": 54.32908419338588,
      "measured_at": null,
      "min_score": 14.61450799184122,
      "model_code_failure_rate": 0.1111111111111111,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": 0.8888888888888888,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 47,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1284,
      "rating_method": "sequential_elo",
      "raw_elo": 1463.6994383303029,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.1111111111111111,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 33.06302955535,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 198,
      "wins": 351
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1043,
      "capped_draws": 52,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 7,
        "cost_basis_counts": {
          "provider_reported": 10
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.1311545862,
        "mean": 0.02955606902857143,
        "median": 0.0063034062,
        "min": 0.0057139812,
        "recorded_attempt_count": 10,
        "run_count": 7,
        "total": 0.2068924832
      },
      "codegen_evaluated_cells": 7,
      "conditional_success_score": 34.32361051437528,
      "contributing_entrants": null,
      "distinct_rated_opponents": 478,
      "draws": 215,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-04"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.5714285714285714,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T14:36:19Z",
      "losses": 469,
      "max_score": 55.17575767805614,
      "measured_at": null,
      "min_score": 11.938705607217443,
      "model_code_failure_rate": 0.14285714285714285,
      "model_code_failures": 1,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": 0.8571428571428571,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 48,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1030,
      "rating_method": "sequential_elo",
      "raw_elo": 1437.439846464674,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.2857142857142857,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 33.02602214581218,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 3,
      "wins": 346
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1228,
      "capped_draws": 162,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 13
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.12956125,
        "mean": 0.031130218749999997,
        "median": 0.0158481,
        "min": 0.0111232,
        "recorded_attempt_count": 13,
        "run_count": 9,
        "total": 0.378603
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 34.8799485598801,
      "contributing_entrants": null,
      "distinct_rated_opponents": 576,
      "draws": 309,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-07"
      ],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.5555555555555556,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:43:56Z",
      "losses": 521,
      "max_score": 63.965994099933845,
      "measured_at": null,
      "min_score": 11.126777189882008,
      "model_code_failure_rate": 0.2222222222222222,
      "model_code_failures": 2,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": 0.7777777777777778,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 49,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1188,
      "rating_method": "sequential_elo",
      "raw_elo": 1389.9450614149937,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.2222222222222222,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 32.75591371384714,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 36,
      "wins": 358
    },
    {
      "admitted_plugins": 25,
      "attempted_matches": 4507,
      "capped_draws": 360,
      "codegen_cost_usd": {
        "admitted_cell_count": 25,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 49
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 7,
        "max": 0.0362005625,
        "mean": 0.024352715625,
        "median": 0.025206018750000003,
        "min": 0.016582062499999998,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 49
        },
        "run_count": 32,
        "token_priced_attempt_count": 49,
        "total": 0.7792869
      },
      "codegen_evaluated_cells": 32,
      "conditional_success_score": 34.0794804876644,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1981,
      "draws": 1101,
      "entry_count": 25,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-08"
      ],
      "first_pass_admitted": 13,
      "first_pass_rate": 0.40625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:14Z",
      "losses": 1856,
      "max_score": 51.745308794671715,
      "measured_at": null,
      "min_score": 12.825845883037182,
      "model_code_failure_rate": 0.21875,
      "model_code_failures": 7,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": 0.78125,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 50,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 4097,
      "rating_method": "sequential_elo",
      "raw_elo": 1397.2986594755037,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 12,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 32.03984993317203,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 402,
      "wins": 1140
    },
    {
      "admitted_plugins": 23,
      "attempted_matches": 3833,
      "capped_draws": 420,
      "codegen_cost_usd": {
        "admitted_cell_count": 23,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 26
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.29749325000000004,
        "mean": 0.24155631250000004,
        "median": 0.24021375,
        "min": 0.17873725,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 26
        },
        "run_count": 24,
        "token_priced_attempt_count": 26,
        "total": 5.7973514999999995
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": 31.71867304290317,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1659,
      "draws": 1313,
      "entry_count": 23,
      "failed_game_aliases": [
        "GB2-06"
      ],
      "first_pass_admitted": 22,
      "first_pass_rate": 0.9166666666666666,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:12Z",
      "losses": 1050,
      "max_score": 41.874942556165,
      "measured_at": null,
      "min_score": 13.393525783812322,
      "model_code_failure_rate": 0.041666666666666664,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": 0.9583333333333334,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 51,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 3327,
      "rating_method": "sequential_elo",
      "raw_elo": 1471.4848087153755,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.041666666666666664,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 31.38297846515569,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 492,
      "wins": 964
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1378,
      "capped_draws": 282,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.156555,
        "mean": 0.1422525,
        "median": 0.14850249999999998,
        "min": 0.12078,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 1.13802
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 31.54347880987911,
      "contributing_entrants": null,
      "distinct_rated_opponents": 571,
      "draws": 590,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-06"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:41Z",
      "losses": 311,
      "max_score": 54.800844026442405,
      "measured_at": null,
      "min_score": 9.392722470573146,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": 0.875,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 52,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1200,
      "rating_method": "sequential_elo",
      "raw_elo": 1482.789410802258,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 30.507849941945324,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 177,
      "wins": 299
    },
    {
      "admitted_plugins": 11,
      "attempted_matches": 2012,
      "capped_draws": 203,
      "codegen_cost_usd": {
        "admitted_cell_count": 11,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 26
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 4,
        "max": 0.015971877,
        "mean": 0.00979196948,
        "median": 0.009500477,
        "min": 0.003562685,
        "recorded_attempt_count": 26,
        "run_count": 15,
        "total": 0.14255357568000002
      },
      "codegen_evaluated_cells": 15,
      "conditional_success_score": 32.263197102394415,
      "contributing_entrants": null,
      "distinct_rated_opponents": 993,
      "draws": 510,
      "entry_count": 11,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-04"
      ],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.3333333333333333,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-08"
      ],
      "infra_failures_excluded": 1,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:54Z",
      "losses": 910,
      "max_score": 60.895193567762064,
      "measured_at": null,
      "min_score": 11.258167918522702,
      "model_code_failure_rate": 0.26666666666666666,
      "model_code_failures": 4,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": 0.7333333333333333,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 53,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 2006,
      "rating_method": "sequential_elo",
      "raw_elo": 1401.1190916032042,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.4,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 29.856078081711996,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 2,
      "wins": 586
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 770,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 4,
        "cell_count": 7,
        "cost_basis_counts": {
          "provider_reported": 15
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.0092087,
        "mean": 0.006571171428571429,
        "median": 0.00666192,
        "min": 0.00456754,
        "recorded_attempt_count": 15,
        "run_count": 7,
        "total": 0.0459982
      },
      "codegen_evaluated_cells": 7,
      "conditional_success_score": 34.3314236075538,
      "contributing_entrants": null,
      "distinct_rated_opponents": 362,
      "draws": 135,
      "entry_count": 4,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-08"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01"
      ],
      "infra_failures_excluded": 1,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:48Z",
      "losses": 379,
      "max_score": 58.56858588745748,
      "measured_at": null,
      "min_score": 19.078764110307976,
      "model_code_failure_rate": 0.42857142857142855,
      "model_code_failures": 3,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": 0.5714285714285714,
      "provider": "Google",
      "provider_key": "google",
      "rank": 54,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 770,
      "rating_method": "sequential_elo",
      "raw_elo": 1391.9757267092311,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 4,
      "repaired_rate": 0.5714285714285714,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 29.8491728115773,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 256
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1389,
      "capped_draws": 171,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 17
        },
        "evidence_basis": "Provider-reported",
        "max": 0.07439860000000001,
        "mean": 0.060660574999999994,
        "median": 0.0565846,
        "min": 0.0485456,
        "recorded_attempt_count": 17,
        "run_count": 8,
        "total": 0.4852846
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 29.208909638100376,
      "contributing_entrants": null,
      "distinct_rated_opponents": 651,
      "draws": 485,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T14:16:21Z",
      "losses": 540,
      "max_score": 63.82915309593504,
      "measured_at": null,
      "min_score": 12.039456124515215,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": 1.0,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 55,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1377,
      "rating_method": "sequential_elo",
      "raw_elo": 1412.3182837229035,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 8,
      "repaired_rate": 1.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 29.208909638100376,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 3,
      "wins": 352
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1069,
      "capped_draws": 148,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 11
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.001632522,
        "mean": 0.00126739675,
        "median": 0.001232095,
        "min": 0.00090937,
        "recorded_attempt_count": 11,
        "run_count": 8,
        "total": 0.010139174
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 30.453962994232025,
      "contributing_entrants": null,
      "distinct_rated_opponents": 465,
      "draws": 418,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-06",
        "GB2-07"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.5,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:05:44Z",
      "losses": 286,
      "max_score": 77.08002467089872,
      "measured_at": null,
      "min_score": 0.0,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": 0.75,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 56,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 984,
      "rating_method": "sequential_elo",
      "raw_elo": 1487.271601952845,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.25,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 28.34060594134785,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 3,
      "wins": 280
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1286,
      "capped_draws": 79,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 13
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.0056201515,
        "mean": 0.0029446039375,
        "median": 0.001956843,
        "min": 0.001266762,
        "recorded_attempt_count": 13,
        "run_count": 9,
        "total": 0.029176983
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 29.722442719806377,
      "contributing_entrants": null,
      "distinct_rated_opponents": 634,
      "draws": 367,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-05"
      ],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.5555555555555556,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-19T20:23:50Z",
      "losses": 485,
      "max_score": 62.30809361425588,
      "measured_at": null,
      "min_score": 8.13506693682872,
      "model_code_failure_rate": 0.2222222222222222,
      "model_code_failures": 2,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": 0.7777777777777778,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 57,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1283,
      "rating_method": "sequential_elo",
      "raw_elo": 1486.7829836576914,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.2222222222222222,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 27.91247720515814,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 2,
      "wins": 431
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 872,
      "capped_draws": 131,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 18
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.01660212,
        "mean": 0.00855959175,
        "median": 0.005979701,
        "min": 0.00271606,
        "recorded_attempt_count": 18,
        "run_count": 8,
        "total": 0.06847673400000001
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 30.64330642131805,
      "contributing_entrants": null,
      "distinct_rated_opponents": 437,
      "draws": 396,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-05"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-13T11:17:15Z",
      "losses": 247,
      "max_score": 75.8950715607423,
      "measured_at": null,
      "min_score": 7.832227043501719,
      "model_code_failure_rate": 0.375,
      "model_code_failures": 3,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": 0.625,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 58,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 871,
      "rating_method": "sequential_elo",
      "raw_elo": 1490.89930790687,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 27.2461804322717,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 228
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 976,
      "capped_draws": 178,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 19
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 4,
        "max": 0.0445032,
        "mean": 0.03204184375,
        "median": 0.0312573,
        "min": 0.0200767,
        "recorded_attempt_count": 19,
        "run_count": 9,
        "total": 0.2854256
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 30.924385903086723,
      "contributing_entrants": null,
      "distinct_rated_opponents": 455,
      "draws": 358,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-04",
        "GB2-05"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:59:37Z",
      "losses": 384,
      "max_score": 62.91507461001133,
      "measured_at": null,
      "min_score": 8.615240570387764,
      "model_code_failure_rate": 0.4444444444444444,
      "model_code_failures": 4,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": 0.5555555555555556,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 59,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 976,
      "rating_method": "sequential_elo",
      "raw_elo": 1419.8732201685204,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.5555555555555556,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 26.69825998643895,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 234
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1006,
      "capped_draws": 133,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 15
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.0,
        "mean": 0.0,
        "median": 0.0,
        "min": 0.0,
        "recorded_attempt_count": 15,
        "run_count": 8,
        "total": 0.0
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 27.523696618442827,
      "contributing_entrants": null,
      "distinct_rated_opponents": 480,
      "draws": 263,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02"
      ],
      "first_pass_admitted": 3,
      "first_pass_rate": 0.375,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:02:27Z",
      "losses": 470,
      "max_score": 48.887746203916265,
      "measured_at": null,
      "min_score": 13.352273105809124,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": 0.75,
      "provider": "Other",
      "provider_key": "other",
      "rank": 60,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 987,
      "rating_method": "sequential_elo",
      "raw_elo": 1380.91863340719,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 25.61368581357492,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 4,
      "wins": 254
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 922,
      "capped_draws": 151,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 18
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.08193244999999999,
        "mean": 0.05495368125,
        "median": 0.056941025,
        "min": 0.02174195,
        "recorded_attempt_count": 18,
        "run_count": 8,
        "total": 0.43962945
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 28.64520773105748,
      "contributing_entrants": null,
      "distinct_rated_opponents": 461,
      "draws": 155,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:05:42Z",
      "losses": 529,
      "max_score": 49.13824204005262,
      "measured_at": null,
      "min_score": 6.917081520365711,
      "model_code_failure_rate": 0.375,
      "model_code_failures": 3,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": 0.625,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 61,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 922,
      "rating_method": "sequential_elo",
      "raw_elo": 1312.4609397125705,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 25.46959155221364,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 238
    },
    {
      "admitted_plugins": 24,
      "attempted_matches": 4122,
      "capped_draws": 465,
      "codegen_cost_usd": {
        "admitted_cell_count": 24,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 29
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.021615250000000003,
        "mean": 0.017288937500000004,
        "median": 0.01830825,
        "min": 0.01259225,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 29
        },
        "run_count": 24,
        "token_priced_attempt_count": 29,
        "total": 0.4149345
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": 25.12416565392437,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1766,
      "draws": 1370,
      "entry_count": 24,
      "failed_game_aliases": [],
      "first_pass_admitted": 19,
      "first_pass_rate": 0.7916666666666666,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:19Z",
      "losses": 1651,
      "max_score": 42.444850252935154,
      "measured_at": null,
      "min_score": 9.614328033579907,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 62,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 3610,
      "rating_method": "sequential_elo",
      "raw_elo": 1357.1138777817043,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.20833333333333337,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 25.12416565392437,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 512,
      "wins": 589
    },
    {
      "admitted_plugins": 25,
      "attempted_matches": 4375,
      "capped_draws": 530,
      "codegen_cost_usd": {
        "admitted_cell_count": 25,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 56
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 7,
        "max": 0.013223525,
        "mean": 0.0089983359375,
        "median": 0.00845865,
        "min": 0.0058062875,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 56
        },
        "run_count": 32,
        "token_priced_attempt_count": 56,
        "total": 0.28794675
      },
      "codegen_evaluated_cells": 32,
      "conditional_success_score": 24.57030339770677,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1929,
      "draws": 1424,
      "entry_count": 25,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07"
      ],
      "first_pass_admitted": 10,
      "first_pass_rate": 0.3125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:48Z",
      "losses": 1571,
      "max_score": 36.40186495059713,
      "measured_at": null,
      "min_score": 7.616710049308469,
      "model_code_failure_rate": 0.21875,
      "model_code_failures": 7,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": 0.78125,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 63,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 3900,
      "rating_method": "sequential_elo",
      "raw_elo": 1401.254992152901,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 15,
      "repaired_rate": 0.46875,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 23.0997897388718,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 449,
      "wins": 905
    },
    {
      "admitted_plugins": 19,
      "attempted_matches": 3412,
      "capped_draws": 288,
      "codegen_cost_usd": {
        "admitted_cell_count": 19,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 34
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 5,
        "max": 0.0027823742,
        "mean": 0.0022230800541666665,
        "median": 0.002112839,
        "min": 0.001825724,
        "recorded_attempt_count": 34,
        "run_count": 24,
        "total": 0.0533539213
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": 23.474400757062273,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1509,
      "draws": 1015,
      "entry_count": 19,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-06"
      ],
      "first_pass_admitted": 14,
      "first_pass_rate": 0.5833333333333334,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:10Z",
      "losses": 1507,
      "max_score": 33.96211011336977,
      "measured_at": null,
      "min_score": 11.9686486964846,
      "model_code_failure_rate": 0.20833333333333337,
      "model_code_failures": 5,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": 0.7916666666666666,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 64,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 3124,
      "rating_method": "sequential_elo",
      "raw_elo": 1353.4368748908826,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.20833333333333337,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 22.142675954211537,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 279,
      "wins": 602
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 876,
      "capped_draws": 208,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 17
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.0597184625,
        "mean": 0.02441709805,
        "median": 0.022888967499999996,
        "min": 0.0040410944,
        "recorded_attempt_count": 17,
        "run_count": 8,
        "total": 0.1953367844
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 24.521755430704992,
      "contributing_entrants": null,
      "distinct_rated_opponents": 439,
      "draws": 487,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-05",
        "GB2-07"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:37:33Z",
      "losses": 264,
      "max_score": 51.50438313656388,
      "measured_at": null,
      "min_score": 11.490618140835323,
      "model_code_failure_rate": 0.375,
      "model_code_failures": 3,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": 0.625,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 65,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 876,
      "rating_method": "sequential_elo",
      "raw_elo": 1415.4539371566143,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 21.80326639021641,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 125
    },
    {
      "admitted_plugins": 14,
      "attempted_matches": 2520,
      "capped_draws": 301,
      "codegen_cost_usd": {
        "admitted_cell_count": 14,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 23
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.004375912,
        "mean": 0.00311882929125,
        "median": 0.002947833025,
        "min": 0.001831717,
        "recorded_attempt_count": 23,
        "run_count": 16,
        "total": 0.04990126866
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": 22.27339080380156,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1070,
      "draws": 770,
      "entry_count": 14,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-04"
      ],
      "first_pass_admitted": 10,
      "first_pass_rate": 0.625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:57Z",
      "losses": 1102,
      "max_score": 39.83364255519466,
      "measured_at": null,
      "min_score": 7.826698848192246,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 2,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": 0.875,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 66,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 2192,
      "rating_method": "sequential_elo",
      "raw_elo": 1339.4856459024777,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 4,
      "repaired_rate": 0.25,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 21.542115517324167,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 321,
      "wins": 320
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 891,
      "capped_draws": 127,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 18
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.0422995,
        "mean": 0.0264053125,
        "median": 0.022682375,
        "min": 0.01946775,
        "recorded_attempt_count": 18,
        "run_count": 8,
        "total": 0.2112425
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 23.913526464397528,
      "contributing_entrants": null,
      "distinct_rated_opponents": 445,
      "draws": 266,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:30Z",
      "losses": 451,
      "max_score": 50.07010386016034,
      "measured_at": null,
      "min_score": 6.198198890963618,
      "model_code_failure_rate": 0.375,
      "model_code_failures": 3,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": 0.625,
      "provider": "Google",
      "provider_key": "google",
      "rank": 67,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 887,
      "rating_method": "sequential_elo",
      "raw_elo": 1314.551923630182,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 21.2624658665295,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 4,
      "wins": 170
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1167,
      "capped_draws": 162,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 7,
        "cost_basis_counts": {
          "provider_reported": 8
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.00391754,
        "mean": 0.0023617857142857143,
        "median": 0.00186618,
        "min": 0.00144036,
        "recorded_attempt_count": 8,
        "run_count": 7,
        "total": 0.0165325
      },
      "codegen_evaluated_cells": 7,
      "conditional_success_score": 21.852753172641428,
      "contributing_entrants": null,
      "distinct_rated_opponents": 548,
      "draws": 288,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.7142857142857143,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-02"
      ],
      "infra_failures_excluded": 1,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:44:35Z",
      "losses": 663,
      "max_score": 38.43379421672025,
      "measured_at": null,
      "min_score": 14.0470230789226,
      "model_code_failure_rate": 0.14285714285714285,
      "model_code_failures": 1,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": 0.8571428571428571,
      "provider": "Google",
      "provider_key": "google",
      "rank": 68,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1167,
      "rating_method": "sequential_elo",
      "raw_elo": 1273.4990604635973,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.14285714285714285,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 21.0266198517886,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 216
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1189,
      "capped_draws": 176,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 16
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.00444719,
        "mean": 0.0024656575,
        "median": 0.002155255,
        "min": 0.001130258,
        "recorded_attempt_count": 16,
        "run_count": 8,
        "total": 0.01972526
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 21.673949582168344,
      "contributing_entrants": null,
      "distinct_rated_opponents": 560,
      "draws": 496,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-07"
      ],
      "first_pass_admitted": 1,
      "first_pass_rate": 0.125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:06:15Z",
      "losses": 482,
      "max_score": 27.909510548557073,
      "measured_at": null,
      "min_score": 14.80141330228754,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": 0.875,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 69,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1173,
      "rating_method": "sequential_elo",
      "raw_elo": 1358.4089995121935,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.75,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 20.962355023916732,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 11,
      "wins": 195
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1346,
      "capped_draws": 144,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 17
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.09705768,
        "mean": 0.0730637925,
        "median": 0.07051098,
        "min": 0.05225028,
        "recorded_attempt_count": 17,
        "run_count": 8,
        "total": 0.58451034
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 20.939560331755704,
      "contributing_entrants": null,
      "distinct_rated_opponents": 627,
      "draws": 426,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-13T11:16:26Z",
      "losses": 675,
      "max_score": 61.310936966259646,
      "measured_at": null,
      "min_score": 3.993926331153074,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": 0.875,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 70,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1346,
      "rating_method": "sequential_elo",
      "raw_elo": 1328.127266670672,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 7,
      "repaired_rate": 0.875,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 20.25207708705361,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 245
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 673,
      "capped_draws": 6,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 17
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.13491350000000002,
        "mean": 0.10247140625,
        "median": 0.113468875,
        "min": 0.061553,
        "recorded_attempt_count": 17,
        "run_count": 8,
        "total": 0.81977125
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 22.564812642546553,
      "contributing_entrants": null,
      "distinct_rated_opponents": 338,
      "draws": 26,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-04"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T12:51:58Z",
      "losses": 488,
      "max_score": 36.75902784191273,
      "measured_at": null,
      "min_score": 2.4688491519502116,
      "model_code_failure_rate": 0.375,
      "model_code_failures": 3,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": 0.625,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 71,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 665,
      "rating_method": "sequential_elo",
      "raw_elo": 1267.7455324044884,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 20.063270856813254,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 4,
      "wins": 151
    },
    {
      "admitted_plugins": 9,
      "attempted_matches": 1785,
      "capped_draws": 316,
      "codegen_cost_usd": {
        "admitted_cell_count": 9,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 33
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 7,
        "max": 0.039679950000000005,
        "mean": 0.029276065625,
        "median": 0.0306386,
        "min": 0.01864685,
        "recorded_attempt_count": 33,
        "run_count": 16,
        "total": 0.46841705
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": 21.843112445435427,
      "contributing_entrants": null,
      "distinct_rated_opponents": 832,
      "draws": 776,
      "entry_count": 9,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-04",
        "GB2-06",
        "GB2-07"
      ],
      "first_pass_admitted": 2,
      "first_pass_rate": 0.125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-19T20:28:27Z",
      "losses": 691,
      "max_score": 36.33021415855389,
      "measured_at": null,
      "min_score": 10.307786108744647,
      "model_code_failure_rate": 0.4375,
      "model_code_failures": 7,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": 0.5625,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 72,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1757,
      "rating_method": "sequential_elo",
      "raw_elo": 1387.8197450002965,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 7,
      "repaired_rate": 0.4375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 18.91669027546711,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 13,
      "wins": 290
    },
    {
      "admitted_plugins": 15,
      "attempted_matches": 2937,
      "capped_draws": 237,
      "codegen_cost_usd": {
        "admitted_cell_count": 15,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 46
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 6,
        "max": 0.02904086,
        "mean": 0.020196066154166664,
        "median": 0.0210709803,
        "min": 0.0085050301,
        "recorded_attempt_count": 46,
        "run_count": 21,
        "total": 0.426437947
      },
      "codegen_evaluated_cells": 21,
      "conditional_success_score": 20.319806819135895,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1351,
      "draws": 1103,
      "entry_count": 15,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-06",
        "GB2-08"
      ],
      "first_pass_admitted": 1,
      "first_pass_rate": 0.047619047619047616,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-03",
        "GB2-06",
        "GB2-07"
      ],
      "infra_failures_excluded": 3,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:44:12Z",
      "losses": 1237,
      "max_score": 31.688441494119953,
      "measured_at": null,
      "min_score": 12.99127610042693,
      "model_code_failure_rate": 0.2857142857142857,
      "model_code_failures": 6,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": 0.7142857142857143,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 73,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 2913,
      "rating_method": "sequential_elo",
      "raw_elo": 1373.8529095474462,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 14,
      "repaired_rate": 0.6666666666666666,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 18.680459977813825,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 24,
      "wins": 573
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1176,
      "capped_draws": 178,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 16
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.0149315,
        "mean": 0.00915140625,
        "median": 0.00943475,
        "min": 0.00402875,
        "recorded_attempt_count": 16,
        "run_count": 8,
        "total": 0.07321125
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 19.47051412152977,
      "contributing_entrants": null,
      "distinct_rated_opponents": 554,
      "draws": 300,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 2,
      "first_pass_rate": 0.25,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:21Z",
      "losses": 677,
      "max_score": 32.62660804381708,
      "measured_at": null,
      "min_score": 7.287827335921403,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": 0.75,
      "provider": "Google",
      "provider_key": "google",
      "rank": 74,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1175,
      "rating_method": "sequential_elo",
      "raw_elo": 1242.0219913838712,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 4,
      "repaired_rate": 0.5,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 18.11935505071165,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 198
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1190,
      "capped_draws": 37,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 13
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.0154064,
        "mean": 0.00874386,
        "median": 0.0064215999999999995,
        "min": 0.00524064,
        "recorded_attempt_count": 13,
        "run_count": 9,
        "total": 0.08145696000000001
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 18.367680108906587,
      "contributing_entrants": null,
      "distinct_rated_opponents": 537,
      "draws": 345,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-08"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.4444444444444444,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-19T23:12:19Z",
      "losses": 636,
      "max_score": 29.091016958055572,
      "measured_at": null,
      "min_score": 7.661044648020888,
      "model_code_failure_rate": 0.3333333333333333,
      "model_code_failures": 3,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": 0.6666666666666666,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 75,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1165,
      "rating_method": "sequential_elo",
      "raw_elo": 1256.219737390031,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.2222222222222222,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 16.597072548072685,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 24,
      "wins": 184
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1368,
      "capped_draws": 164,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.038815,
        "mean": 0.02196775,
        "median": 0.0196975,
        "min": 0.016152,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 0.175742
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 12.591972433354108,
      "contributing_entrants": null,
      "distinct_rated_opponents": 639,
      "draws": 441,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:32Z",
      "losses": 738,
      "max_score": 29.22432622420127,
      "measured_at": null,
      "min_score": 0.0,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 76,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1259,
      "rating_method": "sequential_elo",
      "raw_elo": 1215.330962967301,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 12.591972433354108,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 101,
      "wins": 80
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 1000,
      "capped_draws": 48,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 13
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.00870475,
        "mean": 0.005883593750000001,
        "median": 0.006313125,
        "min": 0.003432,
        "recorded_attempt_count": 13,
        "run_count": 8,
        "total": 0.04706875
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 12.522919483535407,
      "contributing_entrants": null,
      "distinct_rated_opponents": 453,
      "draws": 227,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-04",
        "GB2-08"
      ],
      "first_pass_admitted": 3,
      "first_pass_rate": 0.375,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:43:51Z",
      "losses": 680,
      "max_score": 27.657743183844648,
      "measured_at": null,
      "min_score": 0.0,
      "model_code_failure_rate": 0.375,
      "model_code_failures": 3,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": 0.625,
      "provider": "Google",
      "provider_key": "google",
      "rank": 77,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1000,
      "rating_method": "sequential_elo",
      "raw_elo": 1128.5651786492597,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.25,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 11.13462493557314,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 93
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 856,
      "capped_draws": 8,
      "codegen_cost_usd": {
        "admitted_cell_count": 4,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 4,
        "max": 0.055222,
        "mean": 0.046879,
        "median": 0.046959,
        "min": 0.039474,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 0.375032
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 64.20944368278158,
      "contributing_entrants": null,
      "distinct_rated_opponents": 383,
      "draws": 19,
      "entry_count": 4,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.5,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 24,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:18Z",
      "losses": 332,
      "max_score": 74.07987912466021,
      "measured_at": null,
      "min_score": 37.904090563653,
      "model_code_failure_rate": 0.5,
      "model_code_failures": 4,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": 0.5,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": null,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "provisional",
      "rated_matches": 850,
      "rating_method": "sequential_elo",
      "raw_elo": 1588.2823551530091,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 53.9934910182863,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 6,
      "wins": 499
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1287,
      "capped_draws": 150,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 33
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 9,
        "max": 0.2156775,
        "mean": 0.157461515625,
        "median": 0.15705206249999998,
        "min": 0.098579625,
        "recorded_attempt_count": 33,
        "run_count": 16,
        "total": 2.51938425
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": 36.39955656915741,
      "contributing_entrants": null,
      "distinct_rated_opponents": 572,
      "draws": 433,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:39:29Z",
      "losses": 408,
      "max_score": 74.21628479825006,
      "measured_at": null,
      "min_score": 10.671164514073164,
      "model_code_failure_rate": 0.5625,
      "model_code_failures": 9,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": 0.4375,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": null,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "provisional",
      "rated_matches": 1111,
      "rating_method": "sequential_elo",
      "raw_elo": 1451.722176189691,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 7,
      "repaired_rate": 0.4375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 29.603332785792052,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 148,
      "wins": 270
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 765,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 4,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 17
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 4,
        "max": 0.0711619,
        "mean": 0.0371769625,
        "median": 0.029592175,
        "min": 0.02387735,
        "recorded_attempt_count": 17,
        "run_count": 8,
        "total": 0.2974157
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 27.42111661651889,
      "contributing_entrants": null,
      "distinct_rated_opponents": 358,
      "draws": 307,
      "entry_count": 4,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-07",
        "GB2-08"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:33:38Z",
      "losses": 284,
      "max_score": 45.1976284044549,
      "measured_at": null,
      "min_score": 11.328963024338488,
      "model_code_failure_rate": 0.5,
      "model_code_failures": 4,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": 0.5,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": null,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "provisional",
      "rated_matches": 765,
      "rating_method": "sequential_elo",
      "raw_elo": 1423.2064040998605,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 4,
      "repaired_rate": 0.5,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 23.0583186650848,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 174
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 1041,
      "capped_draws": 244,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 29
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 11,
        "max": 0.0,
        "mean": 0.0,
        "median": 0.0,
        "min": 0.0,
        "recorded_attempt_count": 29,
        "run_count": 16,
        "total": 0.0
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": 23.45410855826144,
      "contributing_entrants": null,
      "distinct_rated_opponents": 463,
      "draws": 244,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-08"
      ],
      "first_pass_admitted": 3,
      "first_pass_rate": 0.1875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:52Z",
      "losses": 568,
      "max_score": 25.92335810041868,
      "measured_at": null,
      "min_score": 20.9848590161042,
      "model_code_failure_rate": 0.6875,
      "model_code_failures": 11,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": 0.3125,
      "provider": "Other",
      "provider_key": "other",
      "rank": null,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "provisional",
      "rated_matches": 957,
      "rating_method": "sequential_elo",
      "raw_elo": 1232.5956088284856,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 17.53603632361322,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 72,
      "wins": 145
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 898,
      "capped_draws": 128,
      "codegen_cost_usd": {
        "admitted_cell_count": 4,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 14
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 4,
        "max": 0.06200699999999999,
        "mean": 0.0429946875,
        "median": 0.0423105,
        "min": 0.0223935,
        "recorded_attempt_count": 14,
        "run_count": 8,
        "total": 0.34395749999999997
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 18.10868187686674,
      "contributing_entrants": null,
      "distinct_rated_opponents": 364,
      "draws": 142,
      "entry_count": 4,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-07"
      ],
      "first_pass_admitted": 2,
      "first_pass_rate": 0.25,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-13T11:16:04Z",
      "losses": 477,
      "max_score": 23.811718728840177,
      "measured_at": null,
      "min_score": 11.269722122355963,
      "model_code_failure_rate": 0.5,
      "model_code_failures": 4,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": 0.5,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": null,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "provisional",
      "rated_matches": 772,
      "rating_method": "sequential_elo",
      "raw_elo": 1213.039686533759,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.25,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 15.227525675227149,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 126,
      "wins": 153
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 727,
      "capped_draws": 32,
      "codegen_cost_usd": {
        "admitted_cell_count": 4,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 13
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 5,
        "max": 0.33185775,
        "mean": 0.18266821875,
        "median": 0.16804049999999998,
        "min": 0.08751,
        "recorded_attempt_count": 13,
        "run_count": 9,
        "total": 1.7932035
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 18.35955162608331,
      "contributing_entrants": null,
      "distinct_rated_opponents": 364,
      "draws": 32,
      "entry_count": 4,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-08"
      ],
      "first_pass_admitted": 2,
      "first_pass_rate": 0.2222222222222222,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-13T11:15:27Z",
      "losses": 577,
      "max_score": 23.80090952267913,
      "measured_at": null,
      "min_score": 10.384314501114984,
      "model_code_failure_rate": 0.5555555555555556,
      "model_code_failures": 5,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": 0.4444444444444444,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": null,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "provisional",
      "rated_matches": 727,
      "rating_method": "sequential_elo",
      "raw_elo": 1128.2233049327087,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.2222222222222222,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 14.990511130063094,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 118
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 201,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 0,
      "max_score": null,
      "measured_at": "2026-07-19T18:25:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 201,
      "rating_method": "sequential_elo",
      "raw_elo": 2217.851148747644,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 100.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 201
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 234,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 203,
      "rating_method": "sequential_elo",
      "raw_elo": 2109.5036475112493,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 91.35944335010576,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 197
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 2098.912558734133,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 90.50823856284876,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 191
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 24,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1968.039713684342,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.05797906734983,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 178
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 271,
      "capped_draws": 22,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 22,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1930.7455809922349,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 77.08002467089872,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 164
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 19,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1895.3679746282792,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 74.0960959782584,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 157
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 42,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 42,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T20:13:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1881.636733207583,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 73.13624142561254,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 145
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 24,
      "max_score": null,
      "measured_at": "2026-07-19T17:08:13Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1871.3032618400064,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 72.31110877690001,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 174
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 34,
      "max_score": null,
      "measured_at": "2026-07-19T19:40:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1863.773097164302,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 71.73223969080041,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 168
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 229,
      "capped_draws": 46,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 46,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:01Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1817.5279719116684,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 68.05592761865948,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 145
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 13,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 27,
      "max_score": null,
      "measured_at": "2026-07-19T22:54:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1816.9440144348723,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 67.83390511372896,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 136
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 4,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 38,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1769.7396233879429,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 64.06461183077373,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 134
    },
    {
      "admitted_plugins": 0,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 0,
      "draws": 20,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 36,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1766.7949819491082,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 63.99399190662281,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 147
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 28,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 28,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 36,
      "max_score": null,
      "measured_at": "2026-07-19T19:31:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1762.925387581977,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 63.67950167428481,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 138
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 209,
      "capped_draws": 47,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 47,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 27,
      "max_score": null,
      "measured_at": "2026-07-13T11:01:25Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1746.3205149839891,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.35359462708452,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 128
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 21,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 21,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 36,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1747.742032182051,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.30809361425588,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 119
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 22,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 22,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-19T20:13:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 201,
      "rating_method": "sequential_elo",
      "raw_elo": 1732.0159014279725,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 61.205822303176355,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 136
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 31,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 31,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 40,
      "max_score": null,
      "measured_at": "2026-07-13T11:01:25Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1722.617410716957,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.47727375069463,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 134
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 41,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 41,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 42,
      "max_score": null,
      "measured_at": "2026-07-13T11:02:25Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1713.1783327478593,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 59.723558844154695,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 122
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 80,
      "capped_draws": 11,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 11,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-18T13:38:53Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1729.0471795844182,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.22785256603646,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 63
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 9,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 58,
      "max_score": null,
      "measured_at": "2026-07-13T11:04:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1689.1285291879904,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 57.80317048612072,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 138
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 209,
      "capped_draws": 36,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 36,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 62,
      "max_score": null,
      "measured_at": "2026-07-18T07:36:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 209,
      "rating_method": "sequential_elo",
      "raw_elo": 1628.8354772367124,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 53.01003732658696,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 111
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 21,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 21,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 60,
      "max_score": null,
      "measured_at": "2026-07-19T22:39:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 178,
      "rating_method": "sequential_elo",
      "raw_elo": 1630.8461985271176,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 52.98735499556829,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 97
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 104,
      "capped_draws": 6,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 6,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 19,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1657.929362886318,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 52.54906076651749,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 24,
      "wins": 55
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 209,
      "capped_draws": 29,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 29,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 62,
      "max_score": null,
      "measured_at": "2026-07-18T07:40:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 209,
      "rating_method": "sequential_elo",
      "raw_elo": 1616.6633385119806,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 52.03808620685494,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 118
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 49,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 49,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 64,
      "max_score": null,
      "measured_at": "2026-07-18T07:45:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 208,
      "rating_method": "sequential_elo",
      "raw_elo": 1577.0187486349464,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.86718004327302,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 95
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 233,
      "capped_draws": 142,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 142,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 209,
      "rating_method": "sequential_elo",
      "raw_elo": 1574.5951623324984,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.67892214642133,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 53
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 48,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 48,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 50,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1571.8654129749273,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.29093722140246,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 82
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 57,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 57,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 44,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:46Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 178,
      "rating_method": "sequential_elo",
      "raw_elo": 1567.8460010354645,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 47.95675898609834,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 77
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 178,
      "capped_draws": 29,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 29,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 73,
      "max_score": null,
      "measured_at": "2026-07-18T12:50:58Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 30,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 178,
      "rating_method": "sequential_elo",
      "raw_elo": 1557.5991820583583,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 47.13854557235755,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 76
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 408,
      "capped_draws": 221,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 183,
      "draws": 221,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 105,
      "max_score": null,
      "measured_at": "2026-07-19T20:19:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 31,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 404,
      "rating_method": "sequential_elo",
      "raw_elo": 1538.7722743262907,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 45.780747588218205,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 78
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 27,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 60,
      "draws": 27,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 51,
      "max_score": null,
      "measured_at": "2026-07-19T20:19:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 32,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 127,
      "rating_method": "sequential_elo",
      "raw_elo": 1555.386633078287,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 45.37208589267169,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 18.508683647302114,
      "unilateral_faults": 1,
      "wins": 49
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 27,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 27,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 75,
      "max_score": null,
      "measured_at": "2026-07-18T07:26:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 33,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 1532.9014110845703,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 45.333741935710094,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 104
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 3,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 3,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 106,
      "max_score": null,
      "measured_at": "2026-07-18T07:26:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 34,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 1532.363238012958,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 45.29076855762162,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 97
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 242,
      "capped_draws": 47,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 47,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 95,
      "max_score": null,
      "measured_at": "2026-07-19T23:11:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 35,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 218,
      "rating_method": "sequential_elo",
      "raw_elo": 1503.843655249275,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 43.07514483050649,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 76
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 15,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 15,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 80,
      "max_score": null,
      "measured_at": "2026-07-19T19:43:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 36,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 1501.9828344124348,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.86487874768688,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 111
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 34,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 34,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 96,
      "max_score": null,
      "measured_at": "2026-07-18T13:13:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 37,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 216,
      "rating_method": "sequential_elo",
      "raw_elo": 1492.7658388863824,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.18065252242629,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 86
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 40,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 40,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 98,
      "max_score": null,
      "measured_at": "2026-07-18T13:20:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 38,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 216,
      "rating_method": "sequential_elo",
      "raw_elo": 1481.1080097227068,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 41.24976926801048,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 78
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 241,
      "capped_draws": 141,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 141,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 35,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:45Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 39,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 217,
      "rating_method": "sequential_elo",
      "raw_elo": 1480.6476034854643,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 41.21798435610656,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 41
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 217,
      "capped_draws": 106,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 106,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 60,
      "max_score": null,
      "measured_at": "2026-07-18T13:26:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 40,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 217,
      "rating_method": "sequential_elo",
      "raw_elo": 1480.5938935798754,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 41.21369559436482,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 51
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 834,
      "capped_draws": 93,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 368,
      "draws": 93,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 384,
      "max_score": null,
      "measured_at": "2026-07-19T23:11:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 41,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 807,
      "rating_method": "sequential_elo",
      "raw_elo": 1465.608239185155,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 39.93290747538863,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 330
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 512,
      "capped_draws": 48,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 225,
      "draws": 48,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 170,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 42,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 461,
      "rating_method": "sequential_elo",
      "raw_elo": 1472.9202679788602,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 39.63678818057228,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 11.445626981434346,
      "unilateral_faults": 51,
      "wins": 243
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 623,
      "capped_draws": 119,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 274,
      "draws": 119,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 255,
      "max_score": null,
      "measured_at": "2026-07-18T13:01:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 43,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 623,
      "rating_method": "sequential_elo",
      "raw_elo": 1452.8751682521522,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 38.95237355194132,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 249
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 217,
      "capped_draws": 33,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 33,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 102,
      "max_score": null,
      "measured_at": "2026-07-18T13:37:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 44,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 217,
      "rating_method": "sequential_elo",
      "raw_elo": 1445.7800593007478,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 38.43379421672025,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 82
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 38,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 38,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 80,
      "max_score": null,
      "measured_at": "2026-07-18T07:22:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 45,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1443.1087668140876,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 38.158374372380486,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 87
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 178,
      "capped_draws": 32,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 32,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 70,
      "max_score": null,
      "measured_at": "2026-07-18T07:07:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 46,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 178,
      "rating_method": "sequential_elo",
      "raw_elo": 1443.5446665239044,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 38.03123776136748,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 76
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 38,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 38,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 80,
      "max_score": null,
      "measured_at": "2026-07-13T09:49:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 47,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1431.1821881901408,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.03063812445185,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 58
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 701,
      "capped_draws": 127,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 316,
      "draws": 127,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 292,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 48,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 641,
      "rating_method": "sequential_elo",
      "raw_elo": 1429.5623071962457,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.40186495059713,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 7.906959036110226,
      "unilateral_faults": 53,
      "wins": 222
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 509,
      "capped_draws": 105,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 222,
      "draws": 105,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 204,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 49,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 457,
      "rating_method": "sequential_elo",
      "raw_elo": 1429.831042124774,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.20768076040539,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 11.13389166034865,
      "unilateral_faults": 52,
      "wins": 148
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 38,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 38,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 97,
      "max_score": null,
      "measured_at": "2026-07-13T09:10:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 50,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 204,
      "rating_method": "sequential_elo",
      "raw_elo": 1418.2850887948575,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.17077061768282,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 69
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 106,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 106,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 68,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 51,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 204,
      "rating_method": "sequential_elo",
      "raw_elo": 1409.0038878685773,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 35.429662268792015,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 30
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 66,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 66,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 84,
      "max_score": null,
      "measured_at": "2026-07-13T10:25:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 52,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1408.3063538759975,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 35.37938499429994,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 55
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 48,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 48,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 92,
      "max_score": null,
      "measured_at": "2026-07-13T10:42:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 53,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1395.7203702322447,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.37438815914073,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 65
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 365,
      "capped_draws": 74,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 183,
      "draws": 74,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 182,
      "max_score": null,
      "measured_at": "2026-07-18T13:38:53Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 54,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 365,
      "rating_method": "sequential_elo",
      "raw_elo": 1393.7558331773266,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.083919098748204,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 109
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 416,
      "capped_draws": 140,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 184,
      "draws": 140,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 156,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 55,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 388,
      "rating_method": "sequential_elo",
      "raw_elo": 1392.0961122348049,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.01864551691267,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 92
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 95,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 95,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 75,
      "max_score": null,
      "measured_at": "2026-07-13T10:33:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 56,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 204,
      "rating_method": "sequential_elo",
      "raw_elo": 1373.9001010765544,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.62660804381708,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 34
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 275,
      "capped_draws": 53,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 128,
      "draws": 53,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 151,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 57,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 246,
      "rating_method": "sequential_elo",
      "raw_elo": 1389.6805588572147,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.54055632230109,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 16.700837490522975,
      "unilateral_faults": 29,
      "wins": 42
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 50,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 50,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 120,
      "max_score": null,
      "measured_at": "2026-07-19T20:13:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 58,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 203,
      "rating_method": "sequential_elo",
      "raw_elo": 1340.0696408267977,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 29.91976859210062,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 33
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 67,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 67,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 90,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 59,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1326.3495764650936,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 28.659703759107938,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 19
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 408,
      "capped_draws": 44,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 180,
      "draws": 44,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 254,
      "max_score": null,
      "measured_at": "2026-07-13T11:05:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 60,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 408,
      "rating_method": "sequential_elo",
      "raw_elo": 1307.2991411331964,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.308489424231603,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 110
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 379,
      "capped_draws": 105,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 179,
      "draws": 105,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 229,
      "max_score": null,
      "measured_at": "2026-07-19T19:43:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 61,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 379,
      "rating_method": "sequential_elo",
      "raw_elo": 1303.5716447340765,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 26.923756518498394,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 45
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 14,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 140,
      "max_score": null,
      "measured_at": "2026-07-13T10:03:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 62,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 204,
      "rating_method": "sequential_elo",
      "raw_elo": 1270.388251226152,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 24.36113714812227,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 50
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 308,
      "capped_draws": 10,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 124,
      "max_score": null,
      "measured_at": "2026-07-13T10:36:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 63,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 182,
      "rating_method": "sequential_elo",
      "raw_elo": 1247.9988882816424,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.44304861230847,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 126,
      "wins": 48
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 43,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 43,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 119,
      "max_score": null,
      "measured_at": "2026-07-13T10:29:49Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 64,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 204,
      "rating_method": "sequential_elo",
      "raw_elo": 1244.695464286948,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.309555778625047,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 42
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 32,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 32,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 141,
      "max_score": null,
      "measured_at": "2026-07-13T10:43:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 65,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 204,
      "rating_method": "sequential_elo",
      "raw_elo": 1228.7411778838975,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.03559835136842,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 31
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 413,
      "capped_draws": 112,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 182,
      "draws": 112,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 219,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 66,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 385,
      "rating_method": "sequential_elo",
      "raw_elo": 1228.9690799556683,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.9848590161042,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 54
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 75,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 75,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 100,
      "max_score": null,
      "measured_at": "2026-07-13T10:53:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 67,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1212.999707894921,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.615415585825644,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 2
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 229,
      "capped_draws": 37,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 37,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 142,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 68,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1199.2151775483187,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.683354201876575,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 26
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 52,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 52,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 131,
      "max_score": null,
      "measured_at": "2026-07-13T10:33:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 69,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 204,
      "rating_method": "sequential_elo",
      "raw_elo": 1193.6662700102047,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.23485012188387,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 21
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 48,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 48,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 145,
      "max_score": null,
      "measured_at": "2026-07-13T10:55:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 70,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1158.8381713068147,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.459230872743149,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 39,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 39,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 118,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:46Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 71,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 175,
      "rating_method": "sequential_elo",
      "raw_elo": 1156.9132723278346,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.123322831265655,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 18
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 863,
      "capped_draws": 141,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 360,
      "draws": 141,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 571,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 72,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 791,
      "rating_method": "sequential_elo",
      "raw_elo": 1126.4084745532457,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.825845883037182,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 79
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 32,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 32,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 131,
      "max_score": null,
      "measured_at": "2026-07-13T11:04:49Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 73,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1097.4794524632837,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.384314501114984,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 13
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 13,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 147,
      "max_score": null,
      "measured_at": "2026-07-13T10:56:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 74,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1054.0578852920023,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 6.917081520365711,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 16
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 83,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T20:22:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1991.6326427339416,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 100.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 164
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 64,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 3,
      "max_score": null,
      "measured_at": "2026-07-19T20:15:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1761.809414029661,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.54693366398842,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 119
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 77,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 3,
      "max_score": null,
      "measured_at": "2026-07-19T19:50:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1741.2859310206684,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 57.03469150252695,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 107
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 78,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 2,
      "max_score": null,
      "measured_at": "2026-07-13T08:41:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1734.0416529644124,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.803509565086735,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 108
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 212,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 126,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-19T22:54:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1628.0873090440657,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.58888031678649,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 25,
      "wins": 57
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 119,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-19T19:25:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1611.4657516747402,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.73354815308172,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 64
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 40,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 23,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 0,
      "max_score": null,
      "measured_at": "2026-07-18T13:32:58Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 40,
      "rating_method": "sequential_elo",
      "raw_elo": 1624.3920370977314,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.234252268553153,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 17
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 89,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1592.071710504841,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.03133313019584,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 67
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 143,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T20:20:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1555.4872213111505,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.090411884130972,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 31
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 83,
      "draws": 132,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 12,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 182,
      "rating_method": "sequential_elo",
      "raw_elo": 1545.289767505208,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.297513470769697,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 4,
      "wins": 38
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 62,
      "draws": 83,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 15,
      "max_score": null,
      "measured_at": "2026-07-19T20:29:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 124,
      "rating_method": "sequential_elo",
      "raw_elo": 1549.741780384856,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.48463931663812,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 19.244078834871203,
      "unilateral_faults": 4,
      "wins": 26
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 158,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 7,
      "max_score": null,
      "measured_at": "2026-07-13T04:24:13Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1526.4347319128728,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.13978047432457,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 23
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 190,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 158,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 20,
      "max_score": null,
      "measured_at": "2026-07-13T03:10:13Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1511.4661034845751,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.581567025013598,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 11
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 195,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 170,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-18T13:32:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 195,
      "rating_method": "sequential_elo",
      "raw_elo": 1501.1789623177865,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.891286352323776,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 9
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 197,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 174,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-18T13:06:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 197,
      "rating_method": "sequential_elo",
      "raw_elo": 1494.6898896066327,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.80141330228754,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 9
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 349,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 165,
      "draws": 301,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 28,
      "max_score": null,
      "measured_at": "2026-07-19T20:20:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 347,
      "rating_method": "sequential_elo",
      "raw_elo": 1496.528840602303,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.791801001532535,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 18
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 220,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 166,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 19,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 195,
      "rating_method": "sequential_elo",
      "raw_elo": 1494.4839505766154,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.741184593734893,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 25,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 194,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 85,
      "draws": 153,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 170,
      "rating_method": "sequential_elo",
      "raw_elo": 1495.7638723484058,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.61450799184122,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 194,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 85,
      "draws": 144,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 12,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:39Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 170,
      "rating_method": "sequential_elo",
      "raw_elo": 1495.6076072588198,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.58766401287794,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 14
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 194,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 85,
      "draws": 161,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 9,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 170,
      "rating_method": "sequential_elo",
      "raw_elo": 1495.1147491406418,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.502998442395736,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 0
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 172,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 144,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-18T13:07:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 172,
      "rating_method": "sequential_elo",
      "raw_elo": 1494.5068949772526,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.429040501509393,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 11
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 197,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 178,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 9,
      "max_score": null,
      "measured_at": "2026-07-18T13:18:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 197,
      "rating_method": "sequential_elo",
      "raw_elo": 1492.2674296240918,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.385271315724854,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 195,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 177,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-18T13:19:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 195,
      "rating_method": "sequential_elo",
      "raw_elo": 1491.0439817266465,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.150249973036471,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 193,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 162,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 22,
      "max_score": null,
      "measured_at": "2026-07-18T13:27:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 193,
      "rating_method": "sequential_elo",
      "raw_elo": 1488.670887417819,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 13.717353446889966,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 9
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 650,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 252,
      "draws": 520,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 42,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 578,
      "rating_method": "sequential_elo",
      "raw_elo": 1486.8152333470407,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 13.393525783812322,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 16
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 151,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-13T03:20:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1487.3409069796846,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 13.07288327052516,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 5
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 571,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 248,
      "draws": 507,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-18T13:29:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 571,
      "rating_method": "sequential_elo",
      "raw_elo": 1484.647053718562,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.99127610042693,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 21
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 164,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-13T03:19:46Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1483.757284789138,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.808440532491185,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 11
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 350,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 165,
      "draws": 303,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 32,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 348,
      "rating_method": "sequential_elo",
      "raw_elo": 1482.030450946625,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.309339079481154,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 13
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 619,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 248,
      "draws": 516,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 30,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 571,
      "rating_method": "sequential_elo",
      "raw_elo": 1480.6048343694315,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.296883952695758,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 179,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-13T03:20:50Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 31,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1480.3814624068068,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.24169582676053,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 2
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 170,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-13T02:44:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 32,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1479.204177726602,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.039456124515215,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 153,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 24,
      "max_score": null,
      "measured_at": "2026-07-13T05:22:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 33,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1478.771633660387,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.938705607217443,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 97,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 54,
      "draws": 83,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 3,
      "max_score": null,
      "measured_at": "2026-07-08T14:06:26Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 34,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 92,
      "rating_method": "sequential_elo",
      "raw_elo": 1490.966094278961,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.919704676069871,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 8.09758473652126,
      "unilateral_faults": 5,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 174,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-13T03:26:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 35,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1478.4258751778857,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.892584996054628,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 211,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 171,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 36,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1476.3888776838344,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.529384168381274,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 162,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 147,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-13T02:43:13Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 37,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1478.3206312567197,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.490618140835323,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 7
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 158,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-13T03:35:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 38,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1476.1966495174229,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.482980464184529,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 15
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 40,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 34,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 2,
      "max_score": null,
      "measured_at": "2026-07-18T13:31:39Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 39,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 40,
      "rating_method": "sequential_elo",
      "raw_elo": 1508.952275545118,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.4034471997158,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 176,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-13T03:18:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 40,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1475.0682309617343,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.328963024338488,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 150,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 41,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1477.1096050255824,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.282582150463083,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 1
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 350,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 165,
      "draws": 306,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-19T20:29:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 42,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 348,
      "rating_method": "sequential_elo",
      "raw_elo": 1475.9113373881587,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.2581679185227,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 151,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 43,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1474.526184085617,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.838789505864996,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 124,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 31,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 44,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1473.5503997432495,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.671164514073164,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 7
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 168,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 15,
      "max_score": null,
      "measured_at": "2026-07-13T03:15:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 45,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1470.2711062713329,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.465061755242804,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 3
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 350,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 165,
      "draws": 312,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 32,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 46,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 348,
      "rating_method": "sequential_elo",
      "raw_elo": 1470.3789422522132,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.307786108744647,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 172,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-13T03:17:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 47,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1468.3676297864029,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.164730563488815,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 172,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-13T02:42:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 48,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1468.472620449504,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.156109092208288,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 1
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 520,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 250,
      "draws": 439,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 53,
      "max_score": null,
      "measured_at": "2026-07-19T20:19:16Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 49,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 519,
      "rating_method": "sequential_elo",
      "raw_elo": 1468.1418423162534,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 9.910010721360797,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 27
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 162,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 145,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-13T03:20:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 50,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1468.7615788904866,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 9.848517519974454,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 6
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 629,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 246,
      "draws": 512,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 41,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:39Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 51,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 557,
      "rating_method": "sequential_elo",
      "raw_elo": 1465.345170469226,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 9.614328033579907,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 211,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 174,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 12,
      "max_score": null,
      "measured_at": "2026-07-19T22:49:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 52,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1463.9508694407311,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 9.392722470573146,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 1
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 162,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 150,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-13T03:30:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 53,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1462.0284427050033,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 8.691866558464978,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 1
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 173,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 15,
      "max_score": null,
      "measured_at": "2026-07-13T03:02:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 54,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1459.2710168841108,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 8.615240570387764,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 1
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 395,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 164,
      "draws": 333,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 36,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 55,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 371,
      "rating_method": "sequential_elo",
      "raw_elo": 1459.5172262340852,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 8.610963018710688,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 2
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 185,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 169,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-13T10:34:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 56,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1458.9800765394484,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 8.511944032346298,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 136,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 23,
      "max_score": null,
      "measured_at": "2026-07-13T03:35:49Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 57,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1458.6915160586964,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 8.13506693682872,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 4
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 565,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 249,
      "draws": 465,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 61,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 58,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 541,
      "rating_method": "sequential_elo",
      "raw_elo": 1455.7451314903628,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.881138090855433,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 15
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 144,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 19,
      "max_score": null,
      "measured_at": "2026-07-13T03:45:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 59,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1456.9286141012149,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.832227043501719,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 430,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 166,
      "draws": 331,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 44,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 60,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 381,
      "rating_method": "sequential_elo",
      "raw_elo": 1454.5728435719077,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.826698848192247,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 166,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 18,
      "max_score": null,
      "measured_at": "2026-07-13T03:32:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 61,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1453.9482658660377,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.661044648020888,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 2
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 791,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 327,
      "draws": 653,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 58,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 62,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 719,
      "rating_method": "sequential_elo",
      "raw_elo": 1454.2298523936438,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.61671004930847,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 163,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 22,
      "max_score": null,
      "measured_at": "2026-07-13T03:49:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 63,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1452.4666259599655,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.433178600457437,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 168,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-13T03:20:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 64,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1444.5772700460002,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 6.05124929753971,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 1
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 160,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 26,
      "max_score": null,
      "measured_at": "2026-07-13T04:41:25Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 65,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1409.3515304162397,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 0.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 51,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 25,
      "draws": 6,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 3,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:08Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 37,
      "rating_method": "sequential_elo",
      "raw_elo": 1726.1253813872777,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 100.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 37.360710565381275,
      "unilateral_faults": 8,
      "wins": 28
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 57,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 24,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 42,
      "rating_method": "sequential_elo",
      "raw_elo": 1656.8215338050852,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 83.07600646134895,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 32.49928516630233,
      "unilateral_faults": 6,
      "wins": 31
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 55,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 3,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1614.322264306601,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 65.27953913368336,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 27,
      "wins": 13
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 60,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 18,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 0,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 26,
      "rating_method": "sequential_elo",
      "raw_elo": 1592.3879631862185,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.27137257662854,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 52.475044864937615,
      "unilateral_faults": 34,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 39,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 21,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 0,
      "max_score": null,
      "measured_at": "2026-07-18T07:22:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 22,
      "rating_method": "sequential_elo",
      "raw_elo": 1592.817950871737,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.72016331680182,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 60.5072070285374,
      "unilateral_faults": 13,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 78,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 36,
      "draws": 25,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 61,
      "rating_method": "sequential_elo",
      "raw_elo": 1559.9612376674422,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.65522539740799,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 19.750063500095248,
      "unilateral_faults": 5,
      "wins": 22
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 142,
      "capped_draws": 14,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 53,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1543.8491195243553,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 53.46761933203254,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 54,
      "wins": 21
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 55,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 6,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 22,
      "rating_method": "sequential_elo",
      "raw_elo": 1560.1206345087253,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 52.26076337429936,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 60.5072070285374,
      "unilateral_faults": 27,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 62,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 23,
      "draws": 21,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 2,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:08Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 28,
      "rating_method": "sequential_elo",
      "raw_elo": 1548.056257492551,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.50097274542412,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 49.09766908852594,
      "unilateral_faults": 33,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 53,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 30,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 39,
      "rating_method": "sequential_elo",
      "raw_elo": 1536.3665196595405,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.33095440558088,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 35.30694150420948,
      "unilateral_faults": 10,
      "wins": 14
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 61,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 2,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 22,
      "rating_method": "sequential_elo",
      "raw_elo": 1544.1967765612917,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.14096707330006,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 60.5072070285374,
      "unilateral_faults": 36,
      "wins": 7
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 31,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 21,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-15T09:06:00Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 21,
      "rating_method": "sequential_elo",
      "raw_elo": 1535.1598787863509,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 45.31798142974239,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 62.850358177805205,
      "unilateral_faults": 3,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 38,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-15T06:47:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1535.4586566733617,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 44.876066552892205,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 9,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 38,
      "capped_draws": 5,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 21,
      "draws": 15,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 3,
      "max_score": null,
      "measured_at": "2026-07-18T15:14:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 23,
      "rating_method": "sequential_elo",
      "raw_elo": 1529.1953708321137,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 44.71416367512211,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 58.31207261596577,
      "unilateral_faults": 7,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 35,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 11,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-15T09:38:53Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 21,
      "rating_method": "sequential_elo",
      "raw_elo": 1527.6395418798847,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 43.37233130702697,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 62.850358177805205,
      "unilateral_faults": 11,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 33,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-15T09:38:53Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1527.705981575098,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.87030624668301,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 4,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 54,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 2,
      "max_score": null,
      "measured_at": "2026-07-19T22:51:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 21,
      "rating_method": "sequential_elo",
      "raw_elo": 1524.8474314032603,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.649960732537416,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 62.850358177805205,
      "unilateral_faults": 30,
      "wins": 6
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 159,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 65,
      "draws": 46,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 69,
      "rating_method": "sequential_elo",
      "raw_elo": 1520.534446289642,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.444850252935154,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 58.45011939260174,
      "unilateral_faults": 90,
      "wins": 13
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 51,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 33,
      "draws": 25,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-18T13:37:53Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 48,
      "rating_method": "sequential_elo",
      "raw_elo": 1487.3909516384024,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 40.238950200339005,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 27.67857142857143,
      "unilateral_faults": 0,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 42,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-18T13:05:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 26,
      "rating_method": "sequential_elo",
      "raw_elo": 1506.35763614118,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 40.01373725333524,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 52.475044864937615,
      "unilateral_faults": 3,
      "wins": 8
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 174,
      "capped_draws": 21,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 64,
      "draws": 40,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 70,
      "rating_method": "sequential_elo",
      "raw_elo": 1507.5695867541574,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 39.17797027985697,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 58.02799666097794,
      "unilateral_faults": 90,
      "wins": 16
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 42,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 17,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-18T16:05:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 26,
      "rating_method": "sequential_elo",
      "raw_elo": 1501.979403510879,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 38.88100756843271,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 52.475044864937615,
      "unilateral_faults": 12,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 61,
      "capped_draws": 6,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 16,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-19T22:51:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 25,
      "rating_method": "sequential_elo",
      "raw_elo": 1499.142262426009,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.76759726733822,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 54.30806756909202,
      "unilateral_faults": 33,
      "wins": 5
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 85,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 41,
      "draws": 20,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 43,
      "rating_method": "sequential_elo",
      "raw_elo": 1502.6034117012573,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.13750839828128,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 61.6787826031713,
      "unilateral_faults": 33,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 30,
      "capped_draws": 5,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 21,
      "draws": 17,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 2,
      "max_score": null,
      "measured_at": "2026-07-18T12:57:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 22,
      "rating_method": "sequential_elo",
      "raw_elo": 1500.245554128479,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.76997367453777,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 60.5072070285374,
      "unilateral_faults": 5,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 31,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 9,
      "max_score": null,
      "measured_at": "2026-07-18T12:51:58Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 24,
      "rating_method": "sequential_elo",
      "raw_elo": 1496.7974806168354,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.75902784191273,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 56.25,
      "unilateral_faults": 3,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 44,
      "capped_draws": 6,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 7,
      "max_score": null,
      "measured_at": "2026-07-18T12:57:26Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 25,
      "rating_method": "sequential_elo",
      "raw_elo": 1493.9724275857818,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.43006546305724,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 54.30806756909202,
      "unilateral_faults": 5,
      "wins": 5
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 122,
      "capped_draws": 7,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 63,
      "draws": 32,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-19T22:54:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 66,
      "rating_method": "sequential_elo",
      "raw_elo": 1491.965718110616,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.59822313047601,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 60.65023878520347,
      "unilateral_faults": 45,
      "wins": 17
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 53,
      "capped_draws": 2,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 27,
      "draws": 20,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 37,
      "rating_method": "sequential_elo",
      "raw_elo": 1472.9634286487662,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.502324757678714,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 37.360710565381275,
      "unilateral_faults": 10,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 51,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 30,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 24,
      "rating_method": "sequential_elo",
      "raw_elo": 1487.841275971651,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.44189220241967,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 56.25,
      "unilateral_faults": 27,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 39,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-18T15:14:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 31,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 24,
      "rating_method": "sequential_elo",
      "raw_elo": 1486.690833169341,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.14425139093331,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 56.25,
      "unilateral_faults": 7,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 39,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 7,
      "max_score": null,
      "measured_at": "2026-07-15T09:56:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 32,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 26,
      "rating_method": "sequential_elo",
      "raw_elo": 1477.2621371920486,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.48619403221036,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 52.475044864937615,
      "unilateral_faults": 3,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 40,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-18T14:16:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 33,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 28,
      "rating_method": "sequential_elo",
      "raw_elo": 1468.8713045965055,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.0143622071467,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 49.09766908852594,
      "unilateral_faults": 3,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 50,
      "capped_draws": 6,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 34,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1479.925857534619,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 30.50870524942116,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 28,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 57,
      "capped_draws": 3,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 35,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1476.7083357203906,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 29.67627290090365,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 30,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 52,
      "capped_draws": 6,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 21,
      "draws": 15,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 36,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 22,
      "rating_method": "sequential_elo",
      "raw_elo": 1471.901085116667,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 29.436735776150815,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 60.5072070285374,
      "unilateral_faults": 29,
      "wins": 2
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 54,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 7,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 37,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1474.9614690906308,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 29.22432622420127,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 29,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 38,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 18,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-14T04:26:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 38,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 23,
      "rating_method": "sequential_elo",
      "raw_elo": 1464.562698958926,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.99249714268673,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 58.31207261596577,
      "unilateral_faults": 4,
      "wins": 1
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 33,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 5,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-15T09:56:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 39,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 21,
      "rating_method": "sequential_elo",
      "raw_elo": 1464.8131949351737,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.11799441187583,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 62.850358177805205,
      "unilateral_faults": 3,
      "wins": 6
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 77,
      "capped_draws": 2,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 26,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:08Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 40,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 34,
      "rating_method": "sequential_elo",
      "raw_elo": 1447.0819507113183,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.101618034124503,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 40.76542547285167,
      "unilateral_faults": 37,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 43,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-14T04:35:01Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 41,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 24,
      "rating_method": "sequential_elo",
      "raw_elo": 1458.2668032208471,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 26.790429608055035,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 56.25,
      "unilateral_faults": 4,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 61,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 17,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 42,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 25,
      "rating_method": "sequential_elo",
      "raw_elo": 1455.028170391493,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 26.35446649301918,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 54.30806756909202,
      "unilateral_faults": 27,
      "wins": 2
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 54,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 43,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1460.6690821362936,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.52662161635416,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 33,
      "wins": 2
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 36,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-15T06:47:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 44,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1439.1506653901242,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.95940957471972,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 11,
      "wins": 1
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 174,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 62,
      "draws": 27,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 35,
      "max_score": null,
      "measured_at": "2026-07-19T23:11:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 45,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 76,
      "rating_method": "sequential_elo",
      "raw_elo": 1424.8120002619244,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.6634480878402,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 53.69706000104055,
      "unilateral_faults": 79,
      "wins": 14
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 57,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 5,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 21,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 46,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 37,
      "rating_method": "sequential_elo",
      "raw_elo": 1379.346598885276,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.281920993460163,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 37.360710565381275,
      "unilateral_faults": 11,
      "wins": 11
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 64,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 25,
      "draws": 5,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 20,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 47,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 35,
      "rating_method": "sequential_elo",
      "raw_elo": 1379.961099285522,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 9.980876737333936,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 39.583333333333336,
      "unilateral_faults": 10,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 56,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 48,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1383.947621293607,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 5.677361994071518,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 28,
      "wins": 1
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 83,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 25,
      "draws": 4,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 22,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 49,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 31,
      "rating_method": "sequential_elo",
      "raw_elo": 1345.4269744064488,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 0.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 44.63834764831843,
      "unilateral_faults": 42,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 23,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 3,
      "max_score": null,
      "measured_at": "2026-07-19T20:25:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1910.0429790696871,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 100.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 160
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 84,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-19T20:11:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1710.7933885474688,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.7273441167311,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 92
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 83,
      "draws": 85,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-19T20:20:09Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1706.679265118354,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 61.943045486748765,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 92
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 95,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-13T07:20:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1612.691053164444,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 44.00470514710234,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 55
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 126,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-19T20:25:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1573.5715975142334,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.04323878185268,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 45
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 107,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 9,
      "max_score": null,
      "measured_at": "2026-07-13T06:51:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1569.6617737816275,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 35.95542625454721,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 47
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 162,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 92,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-13T06:44:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1558.506800704954,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 33.85082376288825,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 54
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 372,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 162,
      "draws": 201,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 46,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 324,
      "rating_method": "sequential_elo",
      "raw_elo": 1548.2603105218825,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.93406247066194,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 77
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 115,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1541.5731286289524,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 30.68312374165008,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 30
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 162,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 97,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 26,
      "max_score": null,
      "measured_at": "2026-07-13T05:55:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1538.5360765579255,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 30.114997119927704,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 39
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 571,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 250,
      "draws": 340,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 49,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 499,
      "rating_method": "sequential_elo",
      "raw_elo": 1526.3772493624135,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.91403849157441,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 110
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 130,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 15,
      "max_score": null,
      "measured_at": "2026-07-13T06:10:45Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1524.2028580866609,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.469379833230263,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 19
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 133,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 18,
      "max_score": null,
      "measured_at": "2026-07-13T05:29:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1518.4707885990597,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 26.39710935603434,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 13
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 100,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 20,
      "max_score": null,
      "measured_at": "2026-07-13T05:07:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1516.882072790436,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 26.082183215449735,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 43
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 112,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 18,
      "max_score": null,
      "measured_at": "2026-07-13T05:24:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1515.9469791946428,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.907259785377427,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 33
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 133,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 18,
      "max_score": null,
      "measured_at": "2026-07-13T05:22:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1514.954893807325,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.72167517689021,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 162,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 122,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 22,
      "max_score": null,
      "measured_at": "2026-07-13T05:17:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1513.9665371375488,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.51889237699729,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 18
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 125,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-19T22:50:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1513.6327304771148,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.456448781739248,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 20
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 106,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 33,
      "max_score": null,
      "measured_at": "2026-07-13T05:27:45Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1512.9440218097536,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.345511091461717,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 24
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 135,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 21,
      "max_score": null,
      "measured_at": "2026-07-13T05:34:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1512.569092518357,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.293107653611095,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 198,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 87,
      "draws": 149,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 174,
      "rating_method": "sequential_elo",
      "raw_elo": 1510.0020478112715,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 24.981785442716276,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 15
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 89,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 25,
      "max_score": null,
      "measured_at": "2026-07-19T20:21:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1507.7468988933458,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 24.744420702191245,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 72
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 83,
      "draws": 129,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 32,
      "max_score": null,
      "measured_at": "2026-07-19T20:09:46Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1507.673579789397,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 24.71601576668925,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 24
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 133,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 25,
      "max_score": null,
      "measured_at": "2026-07-18T13:30:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1506.5058299871805,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 24.35981229662057,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 18
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 130,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1505.2245904025017,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 24.120137012879244,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 30
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 673,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 325,
      "draws": 488,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 79,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 649,
      "rating_method": "sequential_elo",
      "raw_elo": 1505.2434820044568,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.89158662732705,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 82
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 60,
      "draws": 45,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 32,
      "max_score": null,
      "measured_at": "2026-07-19T20:20:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 124,
      "rating_method": "sequential_elo",
      "raw_elo": 1523.7139890557742,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.77593276686914,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 19.244078834871203,
      "unilateral_faults": 4,
      "wins": 47
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 109,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-18T13:32:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1503.0333541118373,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.710233052691397,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 37
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 175,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 142,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 20,
      "max_score": null,
      "measured_at": "2026-07-18T13:22:34Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 175,
      "rating_method": "sequential_elo",
      "raw_elo": 1499.3418928953065,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.003733870007537,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 13
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 174,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 87,
      "draws": 110,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 29,
      "max_score": null,
      "measured_at": "2026-07-18T13:31:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 30,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 174,
      "rating_method": "sequential_elo",
      "raw_elo": 1495.1848218252635,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.209998741054164,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 35
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 127,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 32,
      "max_score": null,
      "measured_at": "2026-07-18T13:30:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 31,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 175,
      "rating_method": "sequential_elo",
      "raw_elo": 1493.0370519838302,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.824317816622944,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 16
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 133,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 28,
      "max_score": null,
      "measured_at": "2026-07-18T13:30:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 32,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 175,
      "rating_method": "sequential_elo",
      "raw_elo": 1492.6536957756052,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.75260522744053,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 14
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 560,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 244,
      "draws": 338,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 87,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:39Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 33,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 488,
      "rating_method": "sequential_elo",
      "raw_elo": 1491.625368506487,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.35151432284935,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 63
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 152,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 18,
      "max_score": null,
      "measured_at": "2026-07-18T13:32:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 34,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1490.1018224099944,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.291194047877852,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 138,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-13T04:51:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 35,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1486.6977379183295,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.45347865004077,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 15
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 135,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-13T04:49:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 36,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1486.637632962378,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.44223510703427,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 199,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 126,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 37,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 175,
      "rating_method": "sequential_elo",
      "raw_elo": 1485.1438466765717,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.347774128566368,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 19
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 107,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 37,
      "max_score": null,
      "measured_at": "2026-07-18T13:30:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 38,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1485.004398644425,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.337643672154105,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 32
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 198,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 87,
      "draws": 133,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 20,
      "max_score": null,
      "measured_at": "2026-07-19T22:39:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 39,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 174,
      "rating_method": "sequential_elo",
      "raw_elo": 1484.1160491905316,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.13941705311589,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 21
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 124,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 23,
      "max_score": null,
      "measured_at": "2026-07-13T04:51:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 40,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1484.8345438109034,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.10493995228072,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 17
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 110,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 29,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 41,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1484.9225058954553,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.085766120064036,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 23
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 122,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 27,
      "max_score": null,
      "measured_at": "2026-07-13T04:50:17Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 42,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1483.2039916831088,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.782187694092556,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 14
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 122,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 23,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 43,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1482.305408162208,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.631826795717306,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 19
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 143,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 44,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1482.1741973198564,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.571653345574234,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 98,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 37,
      "max_score": null,
      "measured_at": "2026-07-13T04:47:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 45,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1481.5462303723896,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.489811083464367,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 29
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 125,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 28,
      "max_score": null,
      "measured_at": "2026-07-13T04:45:16Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 46,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1481.253860215171,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.435118814175468,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 11
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 83,
      "draws": 116,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-19T20:09:46Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 47,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1478.4234125991477,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.24432865630252,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 21
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 348,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 162,
      "draws": 139,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 90,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:16Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 48,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 324,
      "rating_method": "sequential_elo",
      "raw_elo": 1478.8569618553822,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.9511141734381,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 95
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 324,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 162,
      "draws": 241,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-13T09:50:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 49,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 324,
      "rating_method": "sequential_elo",
      "raw_elo": 1476.5573768511383,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.520941944160917,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 35
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 139,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 15,
      "max_score": null,
      "measured_at": "2026-07-13T04:36:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 50,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1475.3273603281289,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.32647718383133,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 162,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 120,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-13T04:47:20Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 51,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1475.4230209775408,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.30874348485692,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 65,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 42,
      "draws": 51,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-08T14:07:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 52,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 58,
      "rating_method": "sequential_elo",
      "raw_elo": 1488.0734362229514,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.7945013339251,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 21.344455490411935,
      "unilateral_faults": 7,
      "wins": 2
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 499,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 250,
      "draws": 353,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 106,
      "max_score": null,
      "measured_at": "2026-07-18T13:32:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 53,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 499,
      "rating_method": "sequential_elo",
      "raw_elo": 1464.180007111608,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.279702862860805,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 40
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 131,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 23,
      "max_score": null,
      "measured_at": "2026-07-19T23:11:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 54,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1463.894901142338,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.152233940860192,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 120,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 29,
      "max_score": null,
      "measured_at": "2026-07-13T05:12:08Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 55,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1462.0402675941834,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.823192329999186,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 14
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 104,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 55,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 56,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1493.2004102426242,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.80923544929635,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 24,
      "wins": 11
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 128,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 25,
      "max_score": null,
      "measured_at": "2026-07-19T22:49:41Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 57,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1461.904495198417,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.779898340965923,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 9
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 118,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 31,
      "max_score": null,
      "measured_at": "2026-07-13T05:16:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 58,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1461.749475059943,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.768795178999987,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 14
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 348,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 164,
      "draws": 226,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 72,
      "max_score": null,
      "measured_at": "2026-07-19T20:20:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 59,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 346,
      "rating_method": "sequential_elo",
      "raw_elo": 1457.4590105074017,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.12805705693795,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 48
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 384,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 170,
      "draws": 151,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 152,
      "max_score": null,
      "measured_at": "2026-07-19T20:21:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 60,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 382,
      "rating_method": "sequential_elo",
      "raw_elo": 1450.4644820831422,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.099220120745768,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 79
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 116,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 29,
      "max_score": null,
      "measured_at": "2026-07-13T06:16:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 61,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1444.813470760353,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.600658879990688,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 18
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 535,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 244,
      "draws": 374,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 89,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 62,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 487,
      "rating_method": "sequential_elo",
      "raw_elo": 1441.4986914364524,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.9686486964846,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 24
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 119,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 29,
      "max_score": null,
      "measured_at": "2026-07-13T05:57:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 63,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1439.4513471639775,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.597592388925714,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 15
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 372,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 162,
      "draws": 213,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 82,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 64,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 324,
      "rating_method": "sequential_elo",
      "raw_elo": 1437.7913644144344,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.269171751144304,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 29
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 80,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 55,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-18T13:27:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 65,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1467.5000467400714,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.001592922177776,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 55,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 65,
      "max_score": null,
      "measured_at": "2026-07-13T06:10:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 66,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1434.8720577315091,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.740966893440644,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 43
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 122,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 29,
      "max_score": null,
      "measured_at": "2026-07-13T05:58:01Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 67,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1432.7007067367829,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.334782779254471,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 121,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 32,
      "max_score": null,
      "measured_at": "2026-07-13T06:02:29Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 68,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1423.787885302052,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 8.667504440488187,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 83,
      "draws": 141,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 37,
      "max_score": null,
      "measured_at": "2026-07-19T20:21:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 69,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1413.940355124856,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.166986972807464,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 214,
      "rating_method": "sequential_elo",
      "raw_elo": 2189.4172617110444,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 100.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 208
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 192,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 2101.5599256868545,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 93.07296360431104,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 187
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 192,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 2007.0964504243177,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 85.75114955106763,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 176
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 245,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 22,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1977.03183596768,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 83.57172066219728,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 199
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 192,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 23,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 1960.8328098869608,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 82.16527921865402,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 169
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 247,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 24,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1944.2879504833952,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 81.04306968418157,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 199
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 28,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 1940.196351882442,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.56575826299661,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 164
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-19T20:24:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 215,
      "rating_method": "sequential_elo",
      "raw_elo": 1933.679334910949,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.18278864235208,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 185
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 213,
      "rating_method": "sequential_elo",
      "raw_elo": 1933.1657077330144,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.13314292967871,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 183
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 408,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 194,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 73,
      "max_score": null,
      "measured_at": "2026-07-19T20:24:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 406,
      "rating_method": "sequential_elo",
      "raw_elo": 1930.4565549318795,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 79.8694620999203,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 333
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 35,
      "max_score": null,
      "measured_at": "2026-07-19T18:36:49Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 216,
      "rating_method": "sequential_elo",
      "raw_elo": 1924.5872917457057,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 79.48293548632148,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 181
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 41,
      "max_score": null,
      "measured_at": "2026-07-13T10:55:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1832.865700872228,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 72.40678999216624,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 182
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 64,
      "max_score": null,
      "measured_at": "2026-07-13T10:52:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1779.9574013908557,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 68.30589565602956,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 159
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1796.7384081258217,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 66.83613208021495,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 48,
      "wins": 72
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 74,
      "max_score": null,
      "measured_at": "2026-07-13T11:06:41Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1756.7107870985228,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 66.50406286570629,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 149
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 1754.0076749994753,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 66.13437229960988,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 144
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 68,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:50Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 214,
      "rating_method": "sequential_elo",
      "raw_elo": 1738.7297749558174,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 65.06745104418093,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 146
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 217,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 77,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:50Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 216,
      "rating_method": "sequential_elo",
      "raw_elo": 1735.1081095036748,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 64.79650419125173,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 139
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 59,
      "max_score": null,
      "measured_at": "2026-07-13T10:54:01Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1724.08563265649,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 63.965994099933845,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 162
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 76,
      "max_score": null,
      "measured_at": "2026-07-13T10:35:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1722.2000400310835,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 63.82915309593504,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 147
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 58,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-19T20:26:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 126,
      "rating_method": "sequential_elo",
      "raw_elo": 1728.4092662022958,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.50648111946364,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 18.75,
      "unilateral_faults": 2,
      "wins": 81
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 54,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 1692.9024546427677,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 61.398138957666426,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 138
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 77,
      "max_score": null,
      "measured_at": "2026-07-13T09:23:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1677.9082554601025,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.39612027754948,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 146
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 73,
      "max_score": null,
      "measured_at": "2026-07-13T10:32:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1665.189160421109,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 59.41026994182745,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 150
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 72,
      "max_score": null,
      "measured_at": "2026-07-13T10:52:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1664.051984424949,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 59.322128031247175,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 151
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 79,
      "max_score": null,
      "measured_at": "2026-07-13T10:41:24Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1659.2010072298715,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.93682106805347,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 142
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 83,
      "max_score": null,
      "measured_at": "2026-07-13T10:00:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1654.3300481477468,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.56858588745748,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 140
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 225,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 100,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 88,
      "max_score": null,
      "measured_at": "2026-07-18T07:30:17Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 225,
      "rating_method": "sequential_elo",
      "raw_elo": 1652.485643981534,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.43481351627299,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 137
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 194,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 82,
      "max_score": null,
      "measured_at": "2026-07-18T07:13:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 194,
      "rating_method": "sequential_elo",
      "raw_elo": 1636.486916757509,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 57.03687454599186,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 112
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 194,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 65,
      "max_score": null,
      "measured_at": "2026-07-18T07:08:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 30,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 194,
      "rating_method": "sequential_elo",
      "raw_elo": 1630.3002526236642,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 56.557349478001406,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 129
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 420,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 198,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 165,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 31,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 419,
      "rating_method": "sequential_elo",
      "raw_elo": 1627.412309265274,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 56.41124513301693,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 254
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 228,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 102,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 99,
      "max_score": null,
      "measured_at": "2026-07-18T07:45:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 32,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 228,
      "rating_method": "sequential_elo",
      "raw_elo": 1619.4103608287369,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.88471911706098,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 129
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 274,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 143,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 80,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 33,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 241,
      "rating_method": "sequential_elo",
      "raw_elo": 1628.9910986264856,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.2402064105688,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 15.813918072220453,
      "unilateral_faults": 29,
      "wins": 161
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 81,
      "max_score": null,
      "measured_at": "2026-07-19T20:15:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 34,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 215,
      "rating_method": "sequential_elo",
      "raw_elo": 1611.284714442574,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.19415196267263,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 134
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 249,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 83,
      "max_score": null,
      "measured_at": "2026-07-19T22:54:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 35,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 225,
      "rating_method": "sequential_elo",
      "raw_elo": 1605.601444995989,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 54.800844026442405,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 142
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 218,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 82,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 36,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 194,
      "rating_method": "sequential_elo",
      "raw_elo": 1601.5519559500542,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 54.32908419338588,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 112
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 196,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 81,
      "max_score": null,
      "measured_at": "2026-07-18T07:45:17Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 37,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 196,
      "rating_method": "sequential_elo",
      "raw_elo": 1589.2056385127698,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 53.383426593216655,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 115
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 198,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 79,
      "max_score": null,
      "measured_at": "2026-07-18T13:28:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 38,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1587.209662701708,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 53.23984773672601,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 119
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 201,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 101,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 85,
      "max_score": null,
      "measured_at": "2026-07-18T12:52:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 39,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 201,
      "rating_method": "sequential_elo",
      "raw_elo": 1564.2836972542352,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.47924961758838,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 116
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 615,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 298,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 254,
      "max_score": null,
      "measured_at": "2026-07-19T20:19:26Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 40,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 609,
      "rating_method": "sequential_elo",
      "raw_elo": 1562.8811661980608,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.37851771163366,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 355
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 233,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 104,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 108,
      "max_score": null,
      "measured_at": "2026-07-18T14:28:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 41,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 233,
      "rating_method": "sequential_elo",
      "raw_elo": 1556.9745692654642,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.067360858185296,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 125
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 255,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 103,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 109,
      "max_score": null,
      "measured_at": "2026-07-19T22:47:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 42,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 231,
      "rating_method": "sequential_elo",
      "raw_elo": 1553.0673307548009,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 50.755795487647504,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 122
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 233,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 104,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 92,
      "max_score": null,
      "measured_at": "2026-07-18T15:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 43,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 233,
      "rating_method": "sequential_elo",
      "raw_elo": 1551.8649996217848,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 50.67132081499249,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 141
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 100,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 107,
      "max_score": null,
      "measured_at": "2026-07-18T13:05:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 44,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 200,
      "rating_method": "sequential_elo",
      "raw_elo": 1534.1507619303593,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 49.13824204005262,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 93
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 229,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 101,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 103,
      "max_score": null,
      "measured_at": "2026-07-18T12:57:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 45,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 229,
      "rating_method": "sequential_elo",
      "raw_elo": 1530.6832284130016,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 49.01198470553688,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 126
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 229,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 102,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 111,
      "max_score": null,
      "measured_at": "2026-07-18T13:02:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 46,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 229,
      "rating_method": "sequential_elo",
      "raw_elo": 1529.080346867643,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.887746203916265,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 118
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 80,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 36,
      "max_score": null,
      "measured_at": "2026-07-18T15:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 47,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1555.026515631915,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.10117101320647,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 44
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 233,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 103,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 130,
      "max_score": null,
      "measured_at": "2026-07-18T13:29:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 48,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 233,
      "rating_method": "sequential_elo",
      "raw_elo": 1493.3178966423554,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 46.133365799256566,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 103
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 235,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 105,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 124,
      "max_score": null,
      "measured_at": "2026-07-18T13:33:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 49,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 235,
      "rating_method": "sequential_elo",
      "raw_elo": 1481.1343053572973,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 45.1976284044549,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 111
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 236,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 106,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 124,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 50,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 208,
      "rating_method": "sequential_elo",
      "raw_elo": 1445.5223847352904,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.31096509029217,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 84
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 225,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 116,
      "max_score": null,
      "measured_at": "2026-07-18T13:05:46Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 51,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 225,
      "rating_method": "sequential_elo",
      "raw_elo": 1432.358546313694,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 41.37287849086039,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 109
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 736,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 334,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 373,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 52,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 683,
      "rating_method": "sequential_elo",
      "raw_elo": 1434.9229171350576,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 40.87409094944158,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 7.906959036110226,
      "unilateral_faults": 53,
      "wins": 310
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 664,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 291,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 395,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 53,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 664,
      "rating_method": "sequential_elo",
      "raw_elo": 1402.9550985696671,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 39.07689302053589,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 269
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 131,
      "max_score": null,
      "measured_at": "2026-07-13T09:11:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 54,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1375.0255732882074,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.91053334109965,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 90
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 543,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 238,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 295,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 55,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 490,
      "rating_method": "sequential_elo",
      "raw_elo": 1383.068198538091,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.67726968699771,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 10.54261204814697,
      "unilateral_faults": 53,
      "wins": 195
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 428,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 202,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 267,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 56,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 421,
      "rating_method": "sequential_elo",
      "raw_elo": 1368.2761811372434,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.33021415855389,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 154
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 384,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 192,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 231,
      "max_score": null,
      "measured_at": "2026-07-13T10:59:39Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 57,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 384,
      "rating_method": "sequential_elo",
      "raw_elo": 1366.2448487733957,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.07908053211064,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 153
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 902,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 400,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 513,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:09Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 58,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 874,
      "rating_method": "sequential_elo",
      "raw_elo": 1359.6807791452063,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 35.70615336320589,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 361
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 147,
      "max_score": null,
      "measured_at": "2026-07-13T09:09:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 59,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1327.4481272619323,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 33.22283063945276,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 74
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 145,
      "max_score": null,
      "measured_at": "2026-07-13T08:46:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 60,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1313.814207923553,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.17538303972932,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 78
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 724,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 298,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 440,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 61,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 676,
      "rating_method": "sequential_elo",
      "raw_elo": 1309.8287024578108,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.87673425923766,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 236
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 446,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 198,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 293,
      "max_score": null,
      "measured_at": "2026-07-13T09:47:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 62,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 446,
      "rating_method": "sequential_elo",
      "raw_elo": 1307.531858998345,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.688441494119953,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 153
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 130,
      "max_score": null,
      "measured_at": "2026-07-13T08:06:26Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 63,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1274.1409172394135,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 29.091016958055572,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 91
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 152,
      "max_score": null,
      "measured_at": "2026-07-13T08:20:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 64,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1263.632103253458,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 28.276484338845368,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 69
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 139,
      "max_score": null,
      "measured_at": "2026-07-13T08:20:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 65,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1258.7774181863872,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.909510548557073,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 84
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 222,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 153,
      "max_score": null,
      "measured_at": "2026-07-13T10:08:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 66,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 222,
      "rating_method": "sequential_elo",
      "raw_elo": 1205.969003299063,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.811718728840177,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 69
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 192,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 147,
      "max_score": null,
      "measured_at": "2026-07-13T10:00:53Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 67,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 1207.836195537406,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.80090952267913,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 45
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 169,
      "max_score": null,
      "measured_at": "2026-07-13T10:39:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 68,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1190.26978308322,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.590209666987988,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 52
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 181,
      "max_score": null,
      "measured_at": "2026-07-13T10:29:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 69,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1168.0026818410374,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.873608699660835,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 42
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 496,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 198,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 350,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 70,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 448,
      "rating_method": "sequential_elo",
      "raw_elo": 1138.2211791864315,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.569853375981936,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 98
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 181,
      "max_score": null,
      "measured_at": "2026-07-13T10:25:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 71,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1122.122660255199,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.317472469475717,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 42
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 739,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 294,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 521,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 72,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 667,
      "rating_method": "sequential_elo",
      "raw_elo": 1109.4597446616813,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.332831857238713,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 146
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 483,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 232,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 392,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 73,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 454,
      "rating_method": "sequential_elo",
      "raw_elo": 1097.923684929318,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.481174430529984,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 11.13389166034865,
      "unilateral_faults": 29,
      "wins": 62
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 186,
      "max_score": null,
      "measured_at": "2026-07-13T11:01:16Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 74,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1080.0485880274048,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.0470230789226,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 35
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 222,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 183,
      "max_score": null,
      "measured_at": "2026-07-13T10:40:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 75,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 222,
      "rating_method": "sequential_elo",
      "raw_elo": 1066.9209979297616,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 13.034181305418958,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 39
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 192,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 164,
      "max_score": null,
      "measured_at": "2026-07-13T10:58:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 76,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 980.732200786807,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 6.198198890963618,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 28
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 198,
      "max_score": null,
      "measured_at": "2026-07-13T10:55:25Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 77,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 967.8087609312524,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 5.347373388052779,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 23
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 210,
      "max_score": null,
      "measured_at": "2026-07-13T10:57:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 78,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 950.2269439894764,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 3.993926331153074,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 13
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 183,
      "max_score": null,
      "measured_at": "2026-07-19T22:54:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 79,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 900.7652140057838,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 0.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 9
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 201,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 200,
      "rating_method": "sequential_elo",
      "raw_elo": 2113.2529956970648,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 100.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 180
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 178,
      "rating_method": "sequential_elo",
      "raw_elo": 2104.6700311392274,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 99.15475058347155,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 164
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 7,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 22,
      "max_score": null,
      "measured_at": "2026-07-19T20:26:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 199,
      "rating_method": "sequential_elo",
      "raw_elo": 2046.7163075917915,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 94.51620294358332,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 170
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 24,
      "max_score": null,
      "measured_at": "2026-07-19T20:21:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 2027.4878813095452,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 92.92727390319158,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 172
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:24Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 2015.6860921297173,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 91.81479306316908,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 153
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 2015.3034273186183,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 91.79025106882118,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 153
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 22,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:08Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 2006.0900644222068,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 91.0317105229074,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 153
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 2000.763555581476,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 90.58621353721448,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 151
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 7,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1965.1257540745448,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 87.65213640227793,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 155
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 25,
      "max_score": null,
      "measured_at": "2026-07-19T20:11:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 199,
      "rating_method": "sequential_elo",
      "raw_elo": 1962.620486155873,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 87.59255399821146,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 162
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 20,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1943.2585453468528,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 85.85876169674393,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 149
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 27,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:50Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 203,
      "rating_method": "sequential_elo",
      "raw_elo": 1941.2653628513624,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 85.85731954540609,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 168
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 19,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:13Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1936.06754569874,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 85.25976029677969,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 143
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 6,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-19T19:49:17Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 200,
      "rating_method": "sequential_elo",
      "raw_elo": 1921.6515965115368,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 84.22536574645959,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 164
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 60,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 124,
      "rating_method": "sequential_elo",
      "raw_elo": 1928.063507325252,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 82.92666095797935,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 19.250031750047626,
      "unilateral_faults": 4,
      "wins": 100
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 226,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 34,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1887.0029000547952,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 81.38419307216813,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 158
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 386,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 182,
      "draws": 20,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 53,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 358,
      "rating_method": "sequential_elo",
      "raw_elo": 1882.2534328321856,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.84990486320305,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 285
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 7,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 34,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1858.78017083648,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 78.90361697494383,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 136
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 5,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 42,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1822.3224384637483,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 75.8950715607423,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 129
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 180,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 35,
      "max_score": null,
      "measured_at": "2026-07-13T11:13:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1811.769918810733,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 75.05378217838995,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 132
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 38,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1798.227461198194,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 73.91132063132736,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 130
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 47,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1770.4153570626615,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 71.76237080818703,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 150
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 37,
      "max_score": null,
      "measured_at": "2026-07-13T11:01:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1758.9621117388845,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 70.67858574947873,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 131
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-13T11:05:20Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1731.3538953086795,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 68.412551151382,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 130
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 185,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 7,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 72,
      "max_score": null,
      "measured_at": "2026-07-18T13:59:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1663.928762676415,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.91507461001133,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 106
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 4,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 23,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1639.618492881366,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.185159863922095,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 48,
      "wins": 53
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 185,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 68,
      "max_score": null,
      "measured_at": "2026-07-18T12:51:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1603.7290733939203,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 57.958805581471466,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 103
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 17,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 84,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 207,
      "rating_method": "sequential_elo",
      "raw_elo": 1592.5756149645515,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 57.171804844224184,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 106
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 259,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 136,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 97,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 230,
      "rating_method": "sequential_elo",
      "raw_elo": 1601.566834229231,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 56.460143022093256,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 16.700837490522975,
      "unilateral_faults": 29,
      "wins": 119
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 634,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 270,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 247,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 30,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 562,
      "rating_method": "sequential_elo",
      "raw_elo": 1570.038337646147,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.19569968343782,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 313
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 376,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 178,
      "draws": 15,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 147,
      "max_score": null,
      "measured_at": "2026-07-19T20:26:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 31,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 373,
      "rating_method": "sequential_elo",
      "raw_elo": 1567.9514463506266,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.02005996423915,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 211
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 190,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 5,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 91,
      "max_score": null,
      "measured_at": "2026-07-18T13:28:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 32,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 190,
      "rating_method": "sequential_elo",
      "raw_elo": 1556.48816333833,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 54.10125469479228,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 94
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 82,
      "max_score": null,
      "measured_at": "2026-07-18T13:24:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 33,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1553.295615896295,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 53.83214821376729,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 97
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 193,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 92,
      "max_score": null,
      "measured_at": "2026-07-18T13:17:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 34,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 193,
      "rating_method": "sequential_elo",
      "raw_elo": 1531.2418555500674,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 52.04121036523053,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 91
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 610,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 286,
      "draws": 16,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 234,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:24Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 35,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 557,
      "rating_method": "sequential_elo",
      "raw_elo": 1547.876116730321,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.745308794671715,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 22.47015523905305,
      "unilateral_faults": 53,
      "wins": 307
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 82,
      "max_score": null,
      "measured_at": "2026-07-18T13:37:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 36,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1525.0221840943957,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.50438313656388,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 97
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 185,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 77,
      "max_score": null,
      "measured_at": "2026-07-18T13:33:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 37,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1524.5024971634189,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.43604390797115,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 94
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 80,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 3,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 35,
      "max_score": null,
      "measured_at": "2026-07-18T13:59:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 38,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1557.562383696878,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.42944139758407,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 42
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 7,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 93,
      "max_score": null,
      "measured_at": "2026-07-18T13:12:49Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 39,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1499.4169506198664,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 49.39629209571456,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 89
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 192,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 93,
      "max_score": null,
      "measured_at": "2026-07-18T13:15:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 40,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 1498.4404656743122,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 49.33453889017528,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 91
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 88,
      "max_score": null,
      "measured_at": "2026-07-18T13:19:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 41,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1492.9147765362331,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.86096501324975,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 89
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 599,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 291,
      "draws": 25,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 264,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 42,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 566,
      "rating_method": "sequential_elo",
      "raw_elo": 1494.4022754125206,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 47.55645299565777,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 19.379453129541623,
      "unilateral_faults": 29,
      "wins": 277
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 213,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 84,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 43,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1471.6275218884668,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 47.10837522132795,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 92
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 190,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 85,
      "max_score": null,
      "measured_at": "2026-07-18T13:32:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 44,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 190,
      "rating_method": "sequential_elo",
      "raw_elo": 1468.277296193658,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 46.83881208054259,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 96
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 181,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 4,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 97,
      "max_score": null,
      "measured_at": "2026-07-18T13:32:41Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 45,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 181,
      "rating_method": "sequential_elo",
      "raw_elo": 1417.0942451711392,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.56668046962967,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 80
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 182,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 110,
      "max_score": null,
      "measured_at": "2026-07-18T13:22:41Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 46,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 182,
      "rating_method": "sequential_elo",
      "raw_elo": 1370.5796066184876,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 38.74378697017428,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 72
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 11,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 109,
      "max_score": null,
      "measured_at": "2026-07-19T20:21:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 47,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 200,
      "rating_method": "sequential_elo",
      "raw_elo": 1359.0254845881047,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.904090563653,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 80
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 566,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 268,
      "draws": 3,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 334,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 48,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 542,
      "rating_method": "sequential_elo",
      "raw_elo": 1358.066580931903,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.704641705021,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 205
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 3,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 108,
      "max_score": null,
      "measured_at": "2026-07-13T11:02:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 49,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1316.3454906374743,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.23774898057906,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 65
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 4,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 110,
      "max_score": null,
      "measured_at": "2026-07-13T10:53:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 50,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1315.3525016186966,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.1559957227449,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 62
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 471,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 213,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 287,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 51,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 418,
      "rating_method": "sequential_elo",
      "raw_elo": 1306.1824899049018,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.50727998813733,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 12.453570188460423,
      "unilateral_faults": 53,
      "wins": 122
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 121,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 52,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1287.166138448974,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.83539904740799,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 53
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 7,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 127,
      "max_score": null,
      "measured_at": "2026-07-19T20:24:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 53,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1284.9657247905943,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.818186314401817,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 68
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 201,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 115,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 54,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1264.153921743513,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 29.947755290899035,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 61
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 130,
      "max_score": null,
      "measured_at": "2026-07-13T10:57:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 55,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1243.9983849389728,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 28.281377385447776,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 46
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 180,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 129,
      "max_score": null,
      "measured_at": "2026-07-13T10:58:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 56,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1236.0895422486844,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.657743183844648,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 49
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 114,
      "max_score": null,
      "measured_at": "2026-07-13T10:53:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 57,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1235.9723508656575,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.627553167229117,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 63
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 135,
      "max_score": null,
      "measured_at": "2026-07-13T10:57:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 58,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1212.42022199375,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.68849520405192,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 41
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 4,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 127,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 59,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1204.7661324053583,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.051367389220953,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 45
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 143,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 60,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1175.6040568173992,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.650439846783147,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 33
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 130,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 61,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1161.8598348623989,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.525834502695627,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 47
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 138,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 62,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1160.555970204923,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.418486705783742,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 37
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 145,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:58Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 63,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1146.417131795777,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.247466438662197,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 31
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 136,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:25Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 64,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1142.6674567245134,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.94571688942271,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 40
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 138,
      "max_score": null,
      "measured_at": "2026-07-13T11:13:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 65,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1132.1373003458395,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.078764110307976,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 37
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 352,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 176,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 286,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:39Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 66,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 352,
      "rating_method": "sequential_elo",
      "raw_elo": 1131.260629418715,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.999624394828466,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 65
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 6,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 131,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 67,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1121.5181057532589,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.204480772726864,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 40
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 132,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 68,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1121.2223397200323,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.180130214648468,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 33
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 405,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 181,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 293,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 69,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 357,
      "rating_method": "sequential_elo",
      "raw_elo": 1115.1927214246716,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.693979671992224,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 62
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 180,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 152,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 70,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1107.073403600676,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.0357831740134,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 28
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 180,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 136,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 71,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1081.256169780604,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.910238043962352,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 30
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 180,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 141,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 72,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1077.5229142525063,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.602877342103772,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 38
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 180,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 149,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 73,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1072.0730196388874,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.154184928501236,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 30
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 142,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 74,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1062.5823635244633,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 13.352273105809124,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 34
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 152,
      "max_score": null,
      "measured_at": "2026-07-19T22:38:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 75,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1046.3479252422885,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.036225680973985,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 26
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 180,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 147,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 76,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1035.3016086951216,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.126777189882008,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 32
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 386,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 179,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 309,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 77,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 362,
      "rating_method": "sequential_elo",
      "raw_elo": 1034.1366678134975,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.03757348862904,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 52
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 168,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 78,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 900.4879497141699,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 0.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 213,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 212,
      "rating_method": "sequential_elo",
      "raw_elo": 2213.60163755207,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 100.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 207
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 214,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 213,
      "rating_method": "sequential_elo",
      "raw_elo": 2191.0952211856857,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 98.42557971783368,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 208
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 207,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 26,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 207,
      "rating_method": "sequential_elo",
      "raw_elo": 1980.0726989542147,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 83.59400879681142,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 181
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 276,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 102,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 50,
      "max_score": null,
      "measured_at": "2026-07-20T08:43:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 275,
      "rating_method": "sequential_elo",
      "raw_elo": 1961.8313657037445,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 82.5709165265804,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 225
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 398,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 184,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 56,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 395,
      "rating_method": "sequential_elo",
      "raw_elo": 1936.8774980083308,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.51515814839607,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 339
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 184,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 28,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1924.4565650704055,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 79.57489937279772,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 156
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 33,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1879.34321498302,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 76.51799749682782,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 172
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 207,
      "rating_method": "sequential_elo",
      "raw_elo": 1866.9421353328428,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 75.65743305436055,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 164
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 215,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-20T08:39:29Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1847.9939468086293,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 74.21628479825006,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 142
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 36,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1829.7985900997423,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 72.93980546270726,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 149
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-20T08:39:26Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 1822.3096281620233,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 72.5215843938676,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 161
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 41,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 207,
      "rating_method": "sequential_elo",
      "raw_elo": 1818.0107769229128,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 72.22469640238779,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 166
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 198,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 47,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1788.9399080122687,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 70.14174273877752,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 151
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 185,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 52,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1744.5797328806148,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 66.95579753782103,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 132
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 207,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 51,
      "max_score": null,
      "measured_at": "2026-07-20T08:39:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 183,
      "rating_method": "sequential_elo",
      "raw_elo": 1744.4392326074358,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 66.94034211955521,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 132
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 59,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 208,
      "rating_method": "sequential_elo",
      "raw_elo": 1724.8419077718663,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 65.69317630293021,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 149
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 183,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 54,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:20Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 183,
      "rating_method": "sequential_elo",
      "raw_elo": 1723.0027444301477,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 65.4364840406668,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 129
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 108,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 42,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 84,
      "rating_method": "sequential_elo",
      "raw_elo": 1748.3221682231233,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 65.0021168384253,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 32.49928516630233,
      "unilateral_faults": 24,
      "wins": 68
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 476,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 226,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 122,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 441,
      "rating_method": "sequential_elo",
      "raw_elo": 1722.9560227490508,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 64.67108012218564,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 12.453570188460423,
      "unilateral_faults": 29,
      "wins": 319
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 210,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-20T08:44:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1708.9329446505674,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 64.46609074562843,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 138
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 207,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 62,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:00Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 207,
      "rating_method": "sequential_elo",
      "raw_elo": 1702.524761583259,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 64.12287600481088,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 145
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 595,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 285,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 205,
      "max_score": null,
      "measured_at": "2026-07-20T08:49:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 585,
      "rating_method": "sequential_elo",
      "raw_elo": 1686.5017799498585,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.93805242641539,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 380
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 209,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 55,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:45Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1683.9317117604437,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.70664219367879,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 130
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 75,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 1665.8468310037947,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 61.54507370770128,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 131
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 90,
      "max_score": null,
      "measured_at": "2026-07-13T10:30:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1662.57676750296,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 61.310936966259646,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 115
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 67,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1658.305456631933,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.914363604458195,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 119
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 375,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 184,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 132,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 371,
      "rating_method": "sequential_elo",
      "raw_elo": 1658.0752540736505,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.895193567762064,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 239
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 183,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 65,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:09Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 183,
      "rating_method": "sequential_elo",
      "raw_elo": 1656.8247006220377,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.79382125920891,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 118
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 620,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 282,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 214,
      "max_score": null,
      "measured_at": "2026-07-20T08:44:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 591,
      "rating_method": "sequential_elo",
      "raw_elo": 1653.2249464492245,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.61574859377792,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 377
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 199,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 67,
      "max_score": null,
      "measured_at": "2026-07-20T08:44:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 30,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 199,
      "rating_method": "sequential_elo",
      "raw_elo": 1648.799998835468,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.31532930836284,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 132
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 79,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 31,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1640.3837262661373,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 59.73960964866883,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 123
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 84,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 42,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 26,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 32,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 84,
      "rating_method": "sequential_elo",
      "raw_elo": 1664.8524239181024,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 59.14636993260279,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 32.49928516630233,
      "unilateral_faults": 0,
      "wins": 58
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 215,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 100,
      "max_score": null,
      "measured_at": "2026-07-18T14:35:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 33,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 215,
      "rating_method": "sequential_elo",
      "raw_elo": 1618.640015606257,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.27438692501393,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 115
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 218,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 83,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:29Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 34,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 194,
      "rating_method": "sequential_elo",
      "raw_elo": 1596.1815838243494,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 56.59865223036674,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 111
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 131,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 58,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 57,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 35,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 127,
      "rating_method": "sequential_elo",
      "raw_elo": 1616.8225752179756,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 56.562076636097146,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 18.508683647302114,
      "unilateral_faults": 4,
      "wins": 70
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 210,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 94,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:41Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 36,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 210,
      "rating_method": "sequential_elo",
      "raw_elo": 1586.3464106719218,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.98636710066346,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 116
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 92,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 37,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 216,
      "rating_method": "sequential_elo",
      "raw_elo": 1580.0717347851164,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.57306742942482,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 124
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 100,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 38,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 216,
      "rating_method": "sequential_elo",
      "raw_elo": 1579.284163468929,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.5178160525783,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 116
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 209,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 89,
      "max_score": null,
      "measured_at": "2026-07-18T14:36:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 39,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 209,
      "rating_method": "sequential_elo",
      "raw_elo": 1574.857203822943,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.17575767805614,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 120
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 289,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 101,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 142,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 40,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 287,
      "rating_method": "sequential_elo",
      "raw_elo": 1560.1947024692745,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 54.43001195791607,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 145
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 419,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 186,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 208,
      "max_score": null,
      "measured_at": "2026-07-20T08:39:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 41,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 417,
      "rating_method": "sequential_elo",
      "raw_elo": 1563.500079681622,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 54.37649620014344,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 209
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 173,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 89,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 42,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 169,
      "rating_method": "sequential_elo",
      "raw_elo": 1505.155510529379,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 50.07010386016034,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 4,
      "wins": 80
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 225,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 118,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 43,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 224,
      "rating_method": "sequential_elo",
      "raw_elo": 1482.9798355750318,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.795842422159346,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 106
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 98,
      "max_score": null,
      "measured_at": "2026-07-12T07:19:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 44,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1458.5825351906751,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 46.891914278195024,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 4,
      "wins": 86
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 113,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:41Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 45,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 199,
      "rating_method": "sequential_elo",
      "raw_elo": 1456.3249209298287,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 46.81240856410825,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 86
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 217,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 119,
      "max_score": null,
      "measured_at": "2026-07-18T13:39:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 46,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 217,
      "rating_method": "sequential_elo",
      "raw_elo": 1453.4926523128145,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 46.69739678054314,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 98
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 201,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 116,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 47,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 201,
      "rating_method": "sequential_elo",
      "raw_elo": 1427.7559875446782,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 44.81802752654815,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 85
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 207,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 123,
      "max_score": null,
      "measured_at": "2026-07-18T07:33:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 48,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 207,
      "rating_method": "sequential_elo",
      "raw_elo": 1398.1757341379953,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.77153617992668,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 84
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 210,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 110,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 49,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1398.9471021180111,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.719305088473725,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 76
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 475,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 204,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 237,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 50,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 422,
      "rating_method": "sequential_elo",
      "raw_elo": 1407.821516156022,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 41.874942556165,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 24.700376732585976,
      "unilateral_faults": 53,
      "wins": 185
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 210,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 110,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 51,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1369.1231225975473,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 40.627029890432105,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 76
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 184,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 120,
      "max_score": null,
      "measured_at": "2026-07-18T13:03:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 52,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1365.4718607095751,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 40.359816348203495,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 64
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 117,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:09Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 53,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1365.238196858895,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 40.343423865130546,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 67
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 182,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 111,
      "max_score": null,
      "measured_at": "2026-07-13T10:08:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 54,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 182,
      "rating_method": "sequential_elo",
      "raw_elo": 1362.3074404228162,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 40.12657589285319,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 71
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 135,
      "max_score": null,
      "measured_at": "2026-07-13T10:26:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 55,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1357.5577991174084,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 39.91259802066606,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 70
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 250,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 107,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 125,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 56,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 201,
      "rating_method": "sequential_elo",
      "raw_elo": 1394.789298445333,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 39.83364255519466,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 42.67472243679027,
      "unilateral_faults": 49,
      "wins": 76
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 425,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 192,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 266,
      "max_score": null,
      "measured_at": "2026-07-20T08:44:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 57,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 425,
      "rating_method": "sequential_elo",
      "raw_elo": 1325.1285742366906,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.67131258724129,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 159
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 198,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 140,
      "max_score": null,
      "measured_at": "2026-07-13T10:08:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 58,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1295.044843928131,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 35.492966497854844,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 58
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 136,
      "max_score": null,
      "measured_at": "2026-07-13T10:26:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 59,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1281.0099027752608,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.54244730456053,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 69
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 619,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 277,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 396,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 60,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 571,
      "rating_method": "sequential_elo",
      "raw_elo": 1273.7950614480271,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 33.96211011336977,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 175
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 140,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 61,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 1262.964769744279,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 33.2812350764154,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 66
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 182,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 122,
      "max_score": null,
      "measured_at": "2026-07-13T10:52:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 62,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 182,
      "rating_method": "sequential_elo",
      "raw_elo": 1259.866347632146,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.9399106271018,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 60
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 209,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 142,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:24Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 63,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1250.712260924394,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.31451053091617,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 43
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 642,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 290,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 417,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 64,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 590,
      "rating_method": "sequential_elo",
      "raw_elo": 1248.2717844825058,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 30.87267632350711,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 21.337361218395134,
      "unilateral_faults": 52,
      "wins": 173
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 398,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 183,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 288,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 65,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 374,
      "rating_method": "sequential_elo",
      "raw_elo": 1194.3117812051496,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 28.368736329667165,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 86
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 199,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 157,
      "max_score": null,
      "measured_at": "2026-07-20T08:44:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 66,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 199,
      "rating_method": "sequential_elo",
      "raw_elo": 1192.6283326358723,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 28.313005051674057,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 42
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 397,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 185,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 284,
      "max_score": null,
      "measured_at": "2026-07-20T08:38:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 67,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 397,
      "rating_method": "sequential_elo",
      "raw_elo": 1180.4988339222468,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.459581015327913,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 113
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 412,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 185,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 300,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 68,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 383,
      "rating_method": "sequential_elo",
      "raw_elo": 1159.1235825951264,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.92335810041868,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 83
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 442,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 198,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 344,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 69,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 389,
      "rating_method": "sequential_elo",
      "raw_elo": 1139.5075927819269,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.935719462016753,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 25.83927083938883,
      "unilateral_faults": 53,
      "wins": 45
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 165,
      "max_score": null,
      "measured_at": "2026-07-13T11:02:13Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 70,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 1109.86793342262,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.540860428737904,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 41
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 183,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 154,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 71,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 183,
      "rating_method": "sequential_elo",
      "raw_elo": 1061.1604399343687,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.005516041876938,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 29
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 182,
      "max_score": null,
      "measured_at": "2026-07-13T11:13:17Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 72,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1019.198467476494,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.175295051675466,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 23
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 184,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 163,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 73,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1010.131975174144,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.431257472598286,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 21
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 198,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 174,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 74,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1003.598767190432,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.046821894435537,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 24
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 211,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 175,
      "max_score": null,
      "measured_at": "2026-07-20T08:43:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 75,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 957.13759060968,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.730010145194916,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 190,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 76,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 892.4461779045539,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.287827335921403,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 16
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 199,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 195,
      "max_score": null,
      "measured_at": "2026-07-20T08:43:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 77,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 199,
      "rating_method": "sequential_elo",
      "raw_elo": 789.0454370328994,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 0.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 2126.6923718578123,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 98.290012518652,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 184
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 213,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 15,
      "max_score": null,
      "measured_at": "2026-07-19T20:20:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 211,
      "rating_method": "sequential_elo",
      "raw_elo": 2043.1402636259547,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 89.47819347812705,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 196
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 21,
      "max_score": null,
      "measured_at": "2026-07-13T11:07:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1963.6270727845756,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.76792349183744,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 168
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-13T11:04:20Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1963.0657653404032,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.70760851946686,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 178
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 19,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 19,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 15,
      "max_score": null,
      "measured_at": "2026-07-13T11:06:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1946.403774774844,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 78.90896540288884,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 154
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 212,
      "capped_draws": 14,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 20,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:26Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1939.7834663198096,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 78.19758392078708,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 154
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 213,
      "capped_draws": 8,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 24,
      "max_score": null,
      "measured_at": "2026-07-19T20:26:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 211,
      "rating_method": "sequential_elo",
      "raw_elo": 1899.8393860063263,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 74.07987912466021,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 179
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 27,
      "max_score": null,
      "measured_at": "2026-07-13T11:01:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1874.9240848051988,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 71.23639433663045,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 162
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 219,
      "capped_draws": 69,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 100,
      "draws": 69,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-19T20:26:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 216,
      "rating_method": "sequential_elo",
      "raw_elo": 1823.0540236330833,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 65.86315907036878,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 131
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 9,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 29,
      "max_score": null,
      "measured_at": "2026-07-13T10:32:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1794.52112484674,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.58850158383014,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 150
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 193,
      "capped_draws": 21,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 21,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-18T13:59:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 193,
      "rating_method": "sequential_elo",
      "raw_elo": 1758.7242311131176,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.78251919520124,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 142
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 80,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 9,
      "max_score": null,
      "measured_at": "2026-07-18T14:36:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1789.3560923707078,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.44733474350692,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 71
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 593,
      "capped_draws": 140,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 285,
      "draws": 140,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 102,
      "max_score": null,
      "measured_at": "2026-07-19T22:47:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 567,
      "rating_method": "sequential_elo",
      "raw_elo": 1730.5062551182962,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.71803693884606,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 325
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 195,
      "capped_draws": 25,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 25,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 58,
      "max_score": null,
      "measured_at": "2026-07-18T13:59:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 195,
      "rating_method": "sequential_elo",
      "raw_elo": 1714.4927144474516,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 54.04543152948865,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 112
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 218,
      "capped_draws": 40,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 100,
      "draws": 40,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 52,
      "max_score": null,
      "measured_at": "2026-07-19T19:54:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 218,
      "rating_method": "sequential_elo",
      "raw_elo": 1670.0153149148105,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 49.43182542421367,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 126
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 80,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-18T15:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1700.4092044504214,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.889597458780386,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 62
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 610,
      "capped_draws": 283,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 286,
      "draws": 283,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 118,
      "max_score": null,
      "measured_at": "2026-07-19T20:26:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 608,
      "rating_method": "sequential_elo",
      "raw_elo": 1635.8068593101143,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 45.64596858389405,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 207
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 196,
      "capped_draws": 32,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 32,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-18T12:51:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 196,
      "rating_method": "sequential_elo",
      "raw_elo": 1607.049252758707,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.50795753707946,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 121
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 220,
      "capped_draws": 52,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 52,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 47,
      "max_score": null,
      "measured_at": "2026-07-19T22:51:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 196,
      "rating_method": "sequential_elo",
      "raw_elo": 1583.3259326638351,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 39.95878166044322,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 97
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 409,
      "capped_draws": 154,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 195,
      "draws": 154,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 94,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 406,
      "rating_method": "sequential_elo",
      "raw_elo": 1568.763898202228,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 38.44569514759158,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 158
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 218,
      "capped_draws": 89,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 89,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 32,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 194,
      "rating_method": "sequential_elo",
      "raw_elo": 1546.4684267251985,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 35.982615143180006,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 73
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 609,
      "capped_draws": 219,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 281,
      "draws": 219,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 161,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 561,
      "rating_method": "sequential_elo",
      "raw_elo": 1526.7555232791372,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 33.8075748539808,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 181
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 78,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 102,
      "draws": 78,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-18T13:18:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1514.225764389215,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.579257820966184,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 79
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 591,
      "capped_draws": 195,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 286,
      "draws": 195,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 199,
      "max_score": null,
      "measured_at": "2026-07-19T20:19:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 587,
      "rating_method": "sequential_elo",
      "raw_elo": 1502.3511242022505,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.251195291498743,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 193
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 104,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 102,
      "draws": 104,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 49,
      "max_score": null,
      "measured_at": "2026-07-18T14:36:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 201,
      "rating_method": "sequential_elo",
      "raw_elo": 1498.275464427157,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 30.85786947221065,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 48
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 89,
      "capped_draws": 23,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 53,
      "draws": 23,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 38,
      "max_score": null,
      "measured_at": "2026-07-12T07:18:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 83,
      "rating_method": "sequential_elo",
      "raw_elo": 1507.320284523039,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 30.164266572203793,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 10.804472558998102,
      "unilateral_faults": 6,
      "wins": 22
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 822,
      "capped_draws": 402,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 377,
      "draws": 402,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 177,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 750,
      "rating_method": "sequential_elo",
      "raw_elo": 1488.4194514279093,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 29.69233519321472,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 171
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 124,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 101,
      "draws": 124,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-18T13:36:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 200,
      "rating_method": "sequential_elo",
      "raw_elo": 1476.4580599067292,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 28.50597915073807,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 31
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 222,
      "capped_draws": 109,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 109,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 47,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1469.93922509029,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.79030487189539,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 42
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 83,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 102,
      "draws": 83,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 72,
      "max_score": null,
      "measured_at": "2026-07-18T13:36:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 30,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1469.102467467738,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.73055959801432,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 47
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 192,
      "capped_draws": 152,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 152,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 23,
      "max_score": null,
      "measured_at": "2026-07-18T13:22:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 31,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 1449.8521929305557,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.584865632593804,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 17
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 194,
      "capped_draws": 58,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 58,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 77,
      "max_score": null,
      "measured_at": "2026-07-18T12:53:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 32,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 194,
      "rating_method": "sequential_elo",
      "raw_elo": 1431.1398224703053,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.590045524889156,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 59
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 212,
      "capped_draws": 140,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 140,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 40,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 33,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1430.026546465291,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.42194383985197,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 8
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 649,
      "capped_draws": 351,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 290,
      "draws": 351,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 146,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 34,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 577,
      "rating_method": "sequential_elo",
      "raw_elo": 1429.1002585284762,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.356176321343483,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 80
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 214,
      "capped_draws": 126,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 126,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 49,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 35,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 211,
      "rating_method": "sequential_elo",
      "raw_elo": 1419.0540128261568,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.41736012440615,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 36
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 154,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 154,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 34,
      "max_score": null,
      "measured_at": "2026-07-18T07:43:08Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 36,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1409.7887856234097,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.247306751579977,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 214,
      "capped_draws": 126,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 126,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 44,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 37,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 190,
      "rating_method": "sequential_elo",
      "raw_elo": 1406.06609545599,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.86369919301597,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 20
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 212,
      "capped_draws": 128,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 128,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 38,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1405.3189584689703,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.76700401531612,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 17
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 218,
      "capped_draws": 146,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 146,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 39,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 191,
      "rating_method": "sequential_elo",
      "raw_elo": 1401.9036805195983,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.42453892869399,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 2
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 138,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 138,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-18T07:37:39Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 40,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1398.8772020896483,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.07480874502237,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 7
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 129,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 129,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-18T07:38:53Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 41,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1398.424496391982,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.026163511594184,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 14
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 132,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 132,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-13T10:37:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 42,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1398.105571299315,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.000131987943323,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 9
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 127,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 127,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-13T11:06:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 43,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1395.8762814960626,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.73567193586623,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 14
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 613,
      "capped_draws": 360,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 284,
      "draws": 360,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 157,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 44,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 565,
      "rating_method": "sequential_elo",
      "raw_elo": 1394.4251675149944,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.599120554232343,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 48
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 109,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 109,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 55,
      "max_score": null,
      "measured_at": "2026-07-13T11:02:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 45,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1394.3186977418932,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.56830265888095,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 22
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 129,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 129,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-13T11:03:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 46,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1387.9420259336252,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.90801399755653,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 12
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 405,
      "capped_draws": 266,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 194,
      "draws": 266,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 108,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 47,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 399,
      "rating_method": "sequential_elo",
      "raw_elo": 1387.0458687407518,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.891337069007747,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 25
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 132,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 132,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 49,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 48,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1386.7927190408395,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.78451587864971,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 132,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 132,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:41Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 49,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1383.3975126241307,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.411447512045243,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 11
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 411,
      "capped_draws": 193,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 194,
      "draws": 193,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 144,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 50,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 387,
      "rating_method": "sequential_elo",
      "raw_elo": 1379.937603115728,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.08326171470216,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 50
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 129,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 129,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 51,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 51,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1379.9491051706354,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.049139240086024,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 9
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 129,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 129,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:26Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 52,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1379.1134741276585,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.93443399794906,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 9
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 430,
      "capped_draws": 248,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 210,
      "draws": 248,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 109,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 53,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 398,
      "rating_method": "sequential_elo",
      "raw_elo": 1396.551451637108,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.591072904112846,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 23.65262231176029,
      "unilateral_faults": 29,
      "wins": 41
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 132,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 132,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 55,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:46Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 54,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1375.6570776322492,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.57139943394984,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 423,
      "capped_draws": 247,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 189,
      "draws": 247,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 110,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 55,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 375,
      "rating_method": "sequential_elo",
      "raw_elo": 1371.9351520062514,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.175614049147185,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 18
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 127,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 127,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 56,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1371.829805973906,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.151771576470264,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 11
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 126,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 126,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 47,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 57,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1370.0230082937594,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.957623158720452,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 13
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 131,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 131,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 46,
      "max_score": null,
      "measured_at": "2026-07-13T11:13:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 58,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1368.1245757862173,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.761998729502363,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 127,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 127,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 53,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 59,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1367.5461615676222,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.691475001584593,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 129,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 129,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 49,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 60,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1363.7606083498813,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.28470053225246,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 8
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 426,
      "capped_draws": 240,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 209,
      "draws": 240,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 131,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 61,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 396,
      "rating_method": "sequential_elo",
      "raw_elo": 1381.1916633843164,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.090109959157065,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 21.786315039332067,
      "unilateral_faults": 30,
      "wins": 25
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 129,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 129,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 49,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:20Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 62,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1361.814201972755,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.075550541321224,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 115,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 115,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 66,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 63,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1361.0557279723228,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.994049148715748,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 126,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 126,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 60,
      "max_score": null,
      "measured_at": "2026-07-13T11:13:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 64,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1357.7971814937282,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.64390388899189,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 212,
      "capped_draws": 127,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 127,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 51,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 65,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1355.313167227012,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.39366014033035,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 127,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 127,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-13T11:13:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 66,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1351.6699661790897,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.985507450020902,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 11
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 132,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 132,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 47,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 67,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1349.190480538579,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.72744628666598,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 8
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 65,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 60,
      "draws": 65,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 51,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 68,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 125,
      "rating_method": "sequential_elo",
      "raw_elo": 1369.618916742115,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.66338227359166,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 18.9970394174356,
      "unilateral_faults": 3,
      "wins": 9
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 133,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 133,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 52,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:45Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 69,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1346.9606253994864,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.479467913302816,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 1
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 125,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 125,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 57,
      "max_score": null,
      "measured_at": "2026-07-13T10:52:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 70,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1341.9363597433005,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 13.95626289544574,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 128,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 128,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 58,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 71,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1336.5754200595743,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 13.36353157637584,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 375,
      "capped_draws": 239,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 189,
      "draws": 239,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 132,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 72,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 375,
      "rating_method": "sequential_elo",
      "raw_elo": 1321.438027318165,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.749474217977998,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 190,
      "capped_draws": 130,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 130,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 56,
      "max_score": null,
      "measured_at": "2026-07-13T11:04:16Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 73,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1319.8723042635704,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.593620745823015,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 118,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 118,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 64,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 74,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1316.934684773728,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.269722122355963,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 212,
      "capped_draws": 29,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 29,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 137,
      "max_score": null,
      "measured_at": "2026-07-19T20:26:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 75,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 210,
      "rating_method": "sequential_elo",
      "raw_elo": 1281.243785441023,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.602053478392984,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 44
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 105,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 105,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 81,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 76,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1263.85170663636,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 5.549046286100777,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 37,
      "capped_draws": 5,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 36,
      "draws": 5,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-08T14:06:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 77,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 36,
      "rating_method": "sequential_elo",
      "raw_elo": 1271.6952570692756,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/mixed"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 2.4688491519502116,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/mixed",
      "track_name": "Reasoning Variants",
      "uncertainty": 38.449493652678655,
      "unilateral_faults": 1,
      "wins": 1
    },
    {
      "admitted_plugins": 15,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 15,
        "cell_count": 16,
        "cost_basis_counts": {
          "local_list_price_estimate": 19
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 1.91725,
        "mean": 1.043635625,
        "median": 0.8756600000000001,
        "min": 0.33027,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 19
        },
        "run_count": 16,
        "token_priced_attempt_count": 19,
        "total": 16.69817
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 15,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 12,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04"
      ],
      "infra_failures_excluded": 4,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:36Z",
      "losses": null,
      "max_score": 98.290012518652,
      "measured_at": null,
      "min_score": 44.00470514710234,
      "model_code_failure_rate": 0.0625,
      "model_code_failures": 1,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1862.978127733007,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.1875,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 79.2245237022498,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 24,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 24,
        "cell_count": 24,
        "cost_basis_counts": {
          "local_standard_estimate": 27
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 2,
        "max": 12.915325,
        "mean": 3.549568958333333,
        "median": 0.4352825,
        "min": 0.09045,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 27
        },
        "run_count": 26,
        "token_priced_attempt_count": 27,
        "total": 85.66335500000001
      },
      "codegen_evaluated_cells": 26,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 24,
      "failed_game_aliases": [
        "GB2-04",
        "GB2-06"
      ],
      "first_pass_admitted": 23,
      "first_pass_rate": 0.8846153846153846,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 24,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:55Z",
      "losses": null,
      "max_score": 100.0,
      "measured_at": null,
      "min_score": 7.602053478392984,
      "model_code_failure_rate": 0.07692307692307693,
      "model_code_failures": 2,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1690.3965732945535,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.038461538461538464,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 57.45182395366876,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 24,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 24,
        "cell_count": 24,
        "cost_basis_counts": {
          "local_standard_estimate": 28
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 3.2933475,
        "mean": 1.078263125,
        "median": 0.5065824999999999,
        "min": 0.06826499999999999,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 28
        },
        "run_count": 24,
        "token_priced_attempt_count": 28,
        "total": 25.878314999999997
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 24,
      "failed_game_aliases": [],
      "first_pass_admitted": 20,
      "first_pass_rate": 0.8333333333333334,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:05Z",
      "losses": null,
      "max_score": 100.0,
      "measured_at": null,
      "min_score": 9.980876737333936,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1715.4549616660672,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 4,
      "repaired_rate": 0.16666666666666666,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 53.790355799083336,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 14,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 14,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 34
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 1.1371809,
        "mean": 0.8062893374999999,
        "median": 0.81347715,
        "min": 0.5199072,
        "recorded_attempt_count": 34,
        "run_count": 16,
        "total": 12.900629399999998
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 14,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-02",
        "GB2-08"
      ],
      "infra_failures_excluded": 2,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:15Z",
      "losses": null,
      "max_score": 82.92666095797935,
      "measured_at": null,
      "min_score": 23.77593276686914,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 2,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1610.2509525723651,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 14,
      "repaired_rate": 0.875,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 52.44822391600324,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 10,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 10,
        "cell_count": 16,
        "cost_basis_counts": {
          "local_standard_estimate": 16
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 6,
        "max": 0.072978,
        "mean": 0.047875499999999994,
        "median": 0.046959,
        "min": 0.03731,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 16
        },
        "run_count": 16,
        "token_priced_attempt_count": 16,
        "total": 0.766008
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 10,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04"
      ],
      "first_pass_admitted": 10,
      "first_pass_rate": 0.625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 24,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:18Z",
      "losses": null,
      "max_score": 74.07987912466021,
      "measured_at": null,
      "min_score": 21.75260522744053,
      "model_code_failure_rate": 0.375,
      "model_code_failures": 6,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1556.5367665636693,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 47.76892991707719,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 22,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 22,
        "cell_count": 24,
        "cost_basis_counts": {
          "local_standard_estimate": 26
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 2,
        "max": 4.4173,
        "mean": 1.10693125,
        "median": 0.4176025,
        "min": 0.12078,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 26
        },
        "run_count": 24,
        "token_priced_attempt_count": 26,
        "total": 26.566350000000003
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 22,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-06"
      ],
      "first_pass_admitted": 21,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:21Z",
      "losses": null,
      "max_score": 100.0,
      "measured_at": null,
      "min_score": 9.392722470573146,
      "model_code_failure_rate": 0.08333333333333333,
      "model_code_failures": 2,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1627.4630268186147,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.041666666666666664,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 47.56960287896285,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 14,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 14,
        "cell_count": 16,
        "cost_basis_counts": {
          "provider_reported": 37
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.6921525,
        "mean": 0.41891503125,
        "median": 0.4031475,
        "min": 0.254883,
        "recorded_attempt_count": 37,
        "run_count": 16,
        "total": 6.702640499999999
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 14,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:20Z",
      "losses": null,
      "max_score": 91.79025106882118,
      "measured_at": null,
      "min_score": 5.549046286100777,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 2,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1655.4435501163562,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 14,
      "repaired_rate": 0.875,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 47.35374141516107,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 22,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 22,
        "cell_count": 24,
        "cost_basis_counts": {
          "local_list_price_estimate": 25
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 2,
        "max": 1.38945,
        "mean": 0.43567666666666666,
        "median": 0.203055,
        "min": 0.131495,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 25
        },
        "run_count": 24,
        "token_priced_attempt_count": 25,
        "total": 10.45624
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 22,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 21,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-04",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 6,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:43:36Z",
      "losses": null,
      "max_score": 91.0317105229074,
      "measured_at": null,
      "min_score": 11.892584996054628,
      "model_code_failure_rate": 0.08333333333333333,
      "model_code_failures": 2,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1654.4839686449968,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.041666666666666664,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 45.752110828723865,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 28,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 28,
        "cell_count": 16,
        "cost_basis_counts": {
          "provider_reported": 56
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 4,
        "max": 0.184451,
        "mean": 0.10049425,
        "median": 0.097957,
        "min": 0.046481,
        "recorded_attempt_count": 56,
        "run_count": 32,
        "total": 3.215816
      },
      "codegen_evaluated_cells": 32,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 28,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 10,
      "first_pass_rate": 0.3125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 16,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:02Z",
      "losses": null,
      "max_score": 66.83613208021495,
      "measured_at": null,
      "min_score": 11.001592922177778,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 4,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1626.2878275277078,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 18,
      "repaired_rate": 0.5625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 45.681358565340275,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 24,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 24,
        "cell_count": 24,
        "cost_basis_counts": {
          "local_standard_estimate": 32
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 12.13752,
        "mean": 2.8005991666666663,
        "median": 0.158095,
        "min": 0.0542775,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 32
        },
        "run_count": 25,
        "token_priced_attempt_count": 32,
        "total": 67.27386999999999
      },
      "codegen_evaluated_cells": 25,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 24,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 17,
      "first_pass_rate": 0.68,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 24,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:22Z",
      "losses": null,
      "max_score": 100.0,
      "measured_at": null,
      "min_score": 10.838789505864996,
      "model_code_failure_rate": 0.04,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1547.9203219709825,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 7,
      "repaired_rate": 0.28,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 42.19352990614933,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 23,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 23,
        "cell_count": 24,
        "cost_basis_counts": {
          "local_list_price_estimate": 25
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.3825,
        "mean": 0.21125,
        "median": 0.22746,
        "min": 0.09594,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 25
        },
        "run_count": 24,
        "token_priced_attempt_count": 25,
        "total": 5.07
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 23,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 22,
      "first_pass_rate": 0.9166666666666666,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:22Z",
      "losses": null,
      "max_score": 78.90896540288884,
      "measured_at": null,
      "min_score": 11.482980464184529,
      "model_code_failure_rate": 0.041666666666666664,
      "model_code_failures": 1,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1557.004813164025,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.041666666666666664,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 42.069798529615646,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 15,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 15,
        "cell_count": 16,
        "cost_basis_counts": {
          "local_standard_estimate": 19
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.454645,
        "mean": 0.22978835937499997,
        "median": 0.1997375,
        "min": 0.1418975,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 19
        },
        "run_count": 16,
        "token_priced_attempt_count": 19,
        "total": 3.67661375
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 15,
      "failed_game_aliases": [
        "GB2-02"
      ],
      "first_pass_admitted": 12,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:07Z",
      "losses": null,
      "max_score": 82.16527921865402,
      "measured_at": null,
      "min_score": 14.429040501509393,
      "model_code_failure_rate": 0.0625,
      "model_code_failures": 1,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1547.374330573245,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.1875,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 41.50372682431928,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 18
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.226689264,
        "mean": 0.16037539425,
        "median": 0.140728509,
        "min": 0.10689655,
        "recorded_attempt_count": 18,
        "run_count": 8,
        "total": 1.283003154
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-06"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T14:35:59Z",
      "losses": null,
      "max_score": 59.41026994182745,
      "measured_at": null,
      "min_score": 17.581567025013598,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1596.5699667703077,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.75,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 40.50672145306643,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 13,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 13,
        "cell_count": 16,
        "cost_basis_counts": {
          "provider_reported": 28
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.10909930499999998,
        "mean": 0.049205097625000006,
        "median": 0.034298199,
        "min": 0.005612805,
        "recorded_attempt_count": 28,
        "run_count": 16,
        "total": 0.787281562
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 13,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04"
      ],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.3125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:09Z",
      "losses": null,
      "max_score": 85.75114955106763,
      "measured_at": null,
      "min_score": 10.465061755242804,
      "model_code_failure_rate": 0.1875,
      "model_code_failures": 3,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1518.0705631281228,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 8,
      "repaired_rate": 0.5,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 38.58539593659148,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 25,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 25,
        "cell_count": 24,
        "cost_basis_counts": {
          "local_list_price_estimate": 27
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.4567,
        "mean": 0.16843436111111112,
        "median": 0.074902,
        "min": 0.049788,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 27
        },
        "run_count": 26,
        "token_priced_attempt_count": 27,
        "total": 4.658466000000001
      },
      "codegen_evaluated_cells": 26,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 23,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 24,
      "first_pass_rate": 0.9230769230769232,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 19,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:21Z",
      "losses": null,
      "max_score": 66.50406286570629,
      "measured_at": null,
      "min_score": 0.0,
      "model_code_failure_rate": 0.038461538461538464,
      "model_code_failures": 1,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1528.7802720140642,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.038461538461538464,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 36.81174373663161,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 12,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 12,
        "cell_count": 15,
        "cost_basis_counts": {
          "provider_reported": 27
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.2996388,
        "mean": 0.08564151856,
        "median": 0.0495450162,
        "min": 0.0057139812,
        "recorded_attempt_count": 27,
        "run_count": 15,
        "total": 1.2846227783999995
      },
      "codegen_evaluated_cells": 15,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 12,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.26666666666666666,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:39:33Z",
      "losses": null,
      "max_score": 68.412551151382,
      "measured_at": null,
      "min_score": 11.938705607217443,
      "model_code_failure_rate": 0.2,
      "model_code_failures": 3,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1474.771973773102,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 8,
      "repaired_rate": 0.5333333333333333,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 36.26958562038245,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 63,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 63,
        "cell_count": 24,
        "cost_basis_counts": {
          "local_list_price_estimate": 1,
          "provider_reported": 150
        },
        "evidence_basis": "Mixed: Provider-reported + Standard list-price estimate",
        "failed_cell_count": 20,
        "max": 0.25293804045,
        "mean": 0.11468783247844444,
        "median": 0.1261480173333333,
        "min": 0.0140238704,
        "recorded_attempt_count": 150,
        "run_count": 83,
        "token_priced_attempt_count": 1,
        "total": 8.90329756963
      },
      "codegen_evaluated_cells": 83,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 62,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "first_pass_admitted": 20,
      "first_pass_rate": 0.24096385542168677,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-03",
        "GB2-06"
      ],
      "infra_failures_excluded": 2,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:49:57Z",
      "losses": null,
      "max_score": 64.67108012218564,
      "measured_at": null,
      "min_score": 9.910010721360797,
      "model_code_failure_rate": 0.24096385542168677,
      "model_code_failures": 20,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1551.231343840457,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 43,
      "repaired_rate": 0.5180722891566265,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 36.11036323157505,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 14,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 14,
        "cell_count": 16,
        "cost_basis_counts": {
          "provider_reported": 24
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.121855,
        "mean": 0.05695140625,
        "median": 0.036751875,
        "min": 0.015285,
        "recorded_attempt_count": 24,
        "run_count": 16,
        "total": 0.9112225
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 14,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-04"
      ],
      "first_pass_admitted": 8,
      "first_pass_rate": 0.5,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:24:06Z",
      "losses": null,
      "max_score": 76.51799749682782,
      "measured_at": null,
      "min_score": 11.919704676069871,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 2,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1505.2911981740397,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.796534608613776,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 24,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 24,
        "cell_count": 24,
        "cost_basis_counts": {
          "local_standard_estimate": 28
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 2,
        "max": 4.987857,
        "mean": 1.074231,
        "median": 0.038529,
        "min": 0.016152,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 28
        },
        "run_count": 26,
        "token_priced_attempt_count": 28,
        "total": 29.297331
      },
      "codegen_evaluated_cells": 26,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 24,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-06"
      ],
      "first_pass_admitted": 22,
      "first_pass_rate": 0.8461538461538461,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 24,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:07Z",
      "losses": null,
      "max_score": 85.25976029677969,
      "measured_at": null,
      "min_score": 0.0,
      "model_code_failure_rate": 0.07692307692307693,
      "model_code_failures": 2,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1488.9611985958566,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.07692307692307693,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.137555986023386,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 59,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 59,
        "cell_count": 23,
        "cost_basis_counts": {
          "local_standard_estimate": 70
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 1.5164726249999998,
        "mean": 0.3999331141304348,
        "median": 0.2331955,
        "min": 0.01259225,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 70
        },
        "run_count": 60,
        "token_priced_attempt_count": 70,
        "total": 18.42633
      },
      "codegen_evaluated_cells": 60,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 59,
      "failed_game_aliases": [
        "GB2-06"
      ],
      "first_pass_admitted": 51,
      "first_pass_rate": 0.85,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-04"
      ],
      "infra_failures_excluded": 6,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:47Z",
      "losses": null,
      "max_score": 80.51515814839607,
      "measured_at": null,
      "min_score": 9.614328033579907,
      "model_code_failure_rate": 0.016666666666666666,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1474.120591693531,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 8,
      "repaired_rate": 0.13333333333333333,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.01581453787501,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 11
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.094594,
        "mean": 0.0571955,
        "median": 0.043012,
        "min": 0.032534,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 11
        },
        "run_count": 8,
        "token_priced_attempt_count": 11,
        "total": 0.457564
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-13T23:05:04Z",
      "losses": null,
      "max_score": 57.80317048612072,
      "measured_at": null,
      "min_score": 8.667504440488187,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1474.9462497663255,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 34.16293015239342,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 65,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 65,
        "cell_count": 24,
        "cost_basis_counts": {
          "local_standard_estimate": 130
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 16,
        "max": 0.38568888333333334,
        "mean": 0.10861519774305556,
        "median": 0.025206018750000003,
        "min": 0.0058062875,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 130
        },
        "run_count": 81,
        "token_priced_attempt_count": 130,
        "total": 6.144237300000001
      },
      "codegen_evaluated_cells": 81,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 65,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "first_pass_admitted": 33,
      "first_pass_rate": 0.4074074074074074,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07"
      ],
      "infra_failures_excluded": 9,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:39:21Z",
      "losses": null,
      "max_score": 51.745308794671715,
      "measured_at": null,
      "min_score": 7.616710049308469,
      "model_code_failure_rate": 0.19753086419753085,
      "model_code_failures": 16,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1466.6886554025066,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 32,
      "repaired_rate": 0.3950617283950617,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 33.126475018100884,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 17
        },
        "evidence_basis": "Provider-reported",
        "max": 0.07439860000000001,
        "mean": 0.060660574999999994,
        "median": 0.0565846,
        "min": 0.0485456,
        "recorded_attempt_count": 17,
        "run_count": 8,
        "total": 0.4852846
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T14:16:21Z",
      "losses": null,
      "max_score": 63.82915309593504,
      "measured_at": null,
      "min_score": 12.039456124515215,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1412.3182837229035,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 8,
      "repaired_rate": 1.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 29.208909638100376,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 13,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 13,
        "cell_count": 16,
        "cost_basis_counts": {
          "provider_reported": 30
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.09803328,
        "mean": 0.049222395,
        "median": 0.05336871,
        "min": 0.00639384,
        "recorded_attempt_count": 30,
        "run_count": 16,
        "total": 0.78755832
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 13,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-04",
        "GB2-06"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.25,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-15T09:38:53Z",
      "losses": null,
      "max_score": 61.310936966259646,
      "measured_at": null,
      "min_score": 3.993926331153074,
      "model_code_failure_rate": 0.1875,
      "model_code_failures": 3,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1406.9181147521276,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 9,
      "repaired_rate": 0.5625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 28.55227506699171,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 13,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 13,
        "cell_count": 16,
        "cost_basis_counts": {
          "provider_reported": 24
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 4,
        "max": 0.0056201515,
        "mean": 0.00210600034375,
        "median": 0.0014932259999999998,
        "min": 0.00090937,
        "recorded_attempt_count": 24,
        "run_count": 17,
        "total": 0.039316157
      },
      "codegen_evaluated_cells": 17,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 13,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-05",
        "GB2-06",
        "GB2-07"
      ],
      "first_pass_admitted": 9,
      "first_pass_rate": 0.5294117647058824,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-19T20:23:50Z",
      "losses": null,
      "max_score": 77.08002467089872,
      "measured_at": null,
      "min_score": 0.0,
      "model_code_failure_rate": 0.2352941176470588,
      "model_code_failures": 4,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1461.1590031868163,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 4,
      "repaired_rate": 0.2352941176470588,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 28.126541573252993,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 12,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 12,
        "cell_count": 16,
        "cost_basis_counts": {
          "provider_reported": 29
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 5,
        "max": 0.08827968,
        "mean": 0.03155999,
        "median": 0.01479536,
        "min": 0.00524064,
        "recorded_attempt_count": 29,
        "run_count": 17,
        "total": 0.51646592
      },
      "codegen_evaluated_cells": 17,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 12,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-04",
        "GB2-08"
      ],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.29411764705882354,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-19T23:12:19Z",
      "losses": null,
      "max_score": 78.90361697494383,
      "measured_at": null,
      "min_score": 7.661044648020888,
      "model_code_failure_rate": 0.29411764705882354,
      "model_code_failures": 5,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1402.694552253526,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 7,
      "repaired_rate": 0.4117647058823529,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 27.55509906029689,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 21,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 21,
        "cell_count": 24,
        "cost_basis_counts": {
          "provider_reported": 65
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 13,
        "max": 0.12956125,
        "mean": 0.030816042708333328,
        "median": 0.0257231,
        "min": 0.0111232,
        "recorded_attempt_count": 65,
        "run_count": 34,
        "total": 1.13244565
      },
      "codegen_evaluated_cells": 34,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 21,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.2058823529411765,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-19T23:12:10Z",
      "losses": null,
      "max_score": 63.965994099933845,
      "measured_at": null,
      "min_score": 8.615240570387764,
      "model_code_failure_rate": 0.38235294117647056,
      "model_code_failures": 13,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1398.2089191127463,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 14,
      "repaired_rate": 0.4117647058823529,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 26.123621325251065,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 10,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 10,
        "cell_count": 14,
        "cost_basis_counts": {
          "provider_reported": 23
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 4,
        "max": 0.0092087,
        "mean": 0.004466478571428571,
        "median": 0.0042425399999999995,
        "min": 0.00144036,
        "recorded_attempt_count": 23,
        "run_count": 14,
        "total": 0.0625307
      },
      "codegen_evaluated_cells": 14,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 10,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-08"
      ],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.35714285714285715,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02"
      ],
      "infra_failures_excluded": 2,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-13T11:05:10Z",
      "losses": null,
      "max_score": 58.56858588745748,
      "measured_at": null,
      "min_score": 14.0470230789226,
      "model_code_failure_rate": 0.2857142857142857,
      "model_code_failures": 4,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1334.8698281057616,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.35714285714285715,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 25.43789633168295,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 37,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 37,
        "cell_count": 24,
        "cost_basis_counts": {
          "provider_reported": 98
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 16,
        "max": 0.0597184625,
        "mean": 0.015795219745833335,
        "median": 0.014928570933333331,
        "min": 0.00169122,
        "recorded_attempt_count": 98,
        "run_count": 53,
        "total": 0.6883146121999998
      },
      "codegen_evaluated_cells": 53,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 37,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "first_pass_admitted": 13,
      "first_pass_rate": 0.24528301886792453,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-03",
        "GB2-06",
        "GB2-07"
      ],
      "infra_failures_excluded": 3,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-19T23:04:38Z",
      "losses": null,
      "max_score": 51.50438313656388,
      "measured_at": null,
      "min_score": 11.490618140835323,
      "model_code_failure_rate": 0.3018867924528302,
      "model_code_failures": 16,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1387.9107916974294,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 24,
      "repaired_rate": 0.4528301886792453,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 25.29173708984834,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 44,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 44,
        "cell_count": 24,
        "cost_basis_counts": {
          "provider_reported": 83
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 11,
        "max": 0.015971877,
        "mean": 0.005044626275138888,
        "median": 0.0029930771,
        "min": 0.001825724,
        "recorded_attempt_count": 83,
        "run_count": 55,
        "total": 0.24580876564000004
      },
      "codegen_evaluated_cells": 55,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 44,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-06"
      ],
      "first_pass_admitted": 29,
      "first_pass_rate": 0.5272727272727272,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-08"
      ],
      "infra_failures_excluded": 1,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:54Z",
      "losses": null,
      "max_score": 60.895193567762064,
      "measured_at": null,
      "min_score": 7.826698848192246,
      "model_code_failure_rate": 0.2,
      "model_code_failures": 11,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 30,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1381.9143038777493,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 15,
      "repaired_rate": 0.2727272727272727,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 24.513623184415906,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 9,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 9,
        "cell_count": 16,
        "cost_basis_counts": {
          "provider_reported": 35
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 7,
        "max": 0.08193244999999999,
        "mean": 0.04606532187500001,
        "median": 0.04204265,
        "min": 0.02174195,
        "recorded_attempt_count": 35,
        "run_count": 16,
        "total": 0.7370451500000001
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 9,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-07",
        "GB2-08"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:32:37Z",
      "losses": null,
      "max_score": 49.13824204005262,
      "measured_at": null,
      "min_score": 6.917081520365711,
      "model_code_failure_rate": 0.4375,
      "model_code_failures": 7,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 31,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1354.8002027874634,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 9,
      "repaired_rate": 0.5625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 24.26395510864922,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 12,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 12,
        "cell_count": 16,
        "cost_basis_counts": {
          "provider_reported": 34
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 4,
        "max": 0.01660212,
        "mean": 0.0055126246249999995,
        "median": 0.0034032865,
        "min": 0.001130258,
        "recorded_attempt_count": 34,
        "run_count": 16,
        "total": 0.088201994
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 12,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-05",
        "GB2-07"
      ],
      "first_pass_admitted": 1,
      "first_pass_rate": 0.0625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T07:43:22Z",
      "losses": null,
      "max_score": 75.8950715607423,
      "measured_at": null,
      "min_score": 7.832227043501719,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 4,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 32,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1391.7681577753738,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 11,
      "repaired_rate": 0.6875,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 24.104267728094214,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 16,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 16,
        "cell_count": 24,
        "cost_basis_counts": {
          "provider_reported": 47
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 8,
        "max": 0.0422995,
        "mean": 0.0138134375,
        "median": 0.00943475,
        "min": 0.003432,
        "recorded_attempt_count": 47,
        "run_count": 24,
        "total": 0.3315225000000001
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 16,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-08"
      ],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.20833333333333337,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:30Z",
      "losses": null,
      "max_score": 50.07010386016034,
      "measured_at": null,
      "min_score": 0.0,
      "model_code_failure_rate": 0.3333333333333333,
      "model_code_failures": 8,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 33,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1275.6434717807404,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 11,
      "repaired_rate": 0.4583333333333333,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 16.83881528427143,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 12,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 12,
        "cell_count": 16,
        "cost_basis_counts": {
          "provider_reported": 50
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 12,
        "max": 0.2156775,
        "mean": 0.12996646093750003,
        "median": 0.12798825,
        "min": 0.061553,
        "recorded_attempt_count": 50,
        "run_count": 24,
        "total": 3.3391555
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 12,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-19T23:12:51Z",
      "losses": null,
      "max_score": 74.21628479825006,
      "measured_at": null,
      "min_score": 2.4688491519502116,
      "model_code_failure_rate": 0.5,
      "model_code_failures": 12,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 34,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1389.7505351723846,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 12,
      "repaired_rate": 0.5,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 24.833301821302655,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 11,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 11,
        "cell_count": 16,
        "cost_basis_counts": {
          "provider_reported": 44
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 13,
        "max": 0.0,
        "mean": 0.0,
        "median": 0.0,
        "min": 0.0,
        "recorded_attempt_count": 44,
        "run_count": 24,
        "total": 0.0
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 11,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-08"
      ],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.25,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-19T23:12:44Z",
      "losses": null,
      "max_score": 48.887746203916265,
      "measured_at": null,
      "min_score": 13.352273105809124,
      "model_code_failure_rate": 0.5416666666666666,
      "model_code_failures": 13,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 35,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1342.93180361881,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.20833333333333337,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 21.574861068594068,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 16,
        "cost_basis_counts": {
          "provider_reported": 27
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 9,
        "max": 0.33185775,
        "mean": 0.112831453125,
        "median": 0.0747585,
        "min": 0.0223935,
        "recorded_attempt_count": 27,
        "run_count": 17,
        "total": 2.137161
      },
      "codegen_evaluated_cells": 17,
      "conditional_success_score": null,
      "contributing_entrants": null,
      "distinct_rated_opponents": 0,
      "draws": null,
      "entry_count": 8,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-07",
        "GB2-08"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.2352941176470588,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-13T11:14:58Z",
      "losses": null,
      "max_score": 23.811718728840177,
      "measured_at": null,
      "min_score": 10.384314501114984,
      "model_code_failure_rate": 0.5294117647058824,
      "model_code_failures": 9,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 36,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1174.3147090573968,
      "reasoning_level": "Overall Models",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "aggregate",
          "source": "public_redacted_snapshot",
          "track_id": "overall"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "aggregate",
        "reasoning_bucket_label": "Aggregate of public reasoning variants",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": null
      },
      "repaired_admitted": 4,
      "repaired_rate": 0.2352941176470588,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 15.109018402645122,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "overall",
      "track_name": "Overall Models",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": null
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1379,
      "capped_draws": 1,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 11
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 1.91725,
        "mean": 1.58262125,
        "median": 1.5854,
        "min": 1.05322,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 11
        },
        "run_count": 8,
        "token_priced_attempt_count": 11,
        "total": 12.66097
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 89.02939366220986,
      "contributing_entrants": null,
      "distinct_rated_opponents": 652,
      "draws": 129,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04"
      ],
      "infra_failures_excluded": 4,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:36Z",
      "losses": 98,
      "max_score": 98.290012518652,
      "measured_at": null,
      "min_score": 72.93980546270726,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": 1.0,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1343,
      "rating_method": "sequential_elo",
      "raw_elo": 1897.6533687768751,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 89.02939366220986,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 35,
      "wins": 1116
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1469,
      "capped_draws": 29,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 12.915325,
        "mean": 10.059243749999998,
        "median": 10.0094025,
        "min": 8.401555,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 80.47395
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 70.61851949173621,
      "contributing_entrants": null,
      "distinct_rated_opponents": 656,
      "draws": 135,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 8,
      "first_pass_rate": 1.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:55Z",
      "losses": 288,
      "max_score": 100.0,
      "measured_at": null,
      "min_score": 7.602053478392984,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1439,
      "rating_method": "sequential_elo",
      "raw_elo": 1778.6799910644754,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 70.61851949173621,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 15,
      "wins": 1016
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1314,
      "capped_draws": 111,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 4.4173,
        "mean": 2.7356843750000004,
        "median": 2.536595,
        "min": 2.045485,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 21.885474999999996
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 69.42492057631782,
      "contributing_entrants": null,
      "distinct_rated_opponents": 590,
      "draws": 264,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-02"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:21Z",
      "losses": 154,
      "max_score": 100.0,
      "measured_at": null,
      "min_score": 19.24432865630252,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": 0.875,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1287,
      "rating_method": "sequential_elo",
      "raw_elo": 1834.840688937272,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 67.14557617248093,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 12,
      "wins": 869
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1488,
      "capped_draws": 44,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 3.2933475,
        "mean": 2.642580625,
        "median": 2.681945,
        "min": 1.562975,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 21.140645
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 64.66453134183178,
      "contributing_entrants": null,
      "distinct_rated_opponents": 666,
      "draws": 263,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:05Z",
      "losses": 122,
      "max_score": 100.0,
      "measured_at": null,
      "min_score": 9.980876737333936,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1454,
      "rating_method": "sequential_elo",
      "raw_elo": 1860.100711513732,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 64.66453134183178,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 15,
      "wins": 1069
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1395,
      "capped_draws": 70,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 11
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 12.13752,
        "mean": 8.1660015625,
        "median": 8.18098,
        "min": 4.7142525,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 11
        },
        "run_count": 8,
        "token_priced_attempt_count": 11,
        "total": 65.32801249999999
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 57.60709859351452,
      "contributing_entrants": null,
      "distinct_rated_opponents": 656,
      "draws": 301,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:22Z",
      "losses": 272,
      "max_score": 100.0,
      "measured_at": null,
      "min_score": 23.297513470769697,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1373,
      "rating_method": "sequential_elo",
      "raw_elo": 1673.782357430351,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 57.60709859351452,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 16,
      "wins": 800
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1307,
      "capped_draws": 11,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 10
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 4.987857,
        "mean": 3.1645696249999995,
        "median": 3.0710854999999997,
        "min": 1.985613,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 10
        },
        "run_count": 9,
        "token_priced_attempt_count": 10,
        "total": 28.806541000000003
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 54.66589143146764,
      "contributing_entrants": null,
      "distinct_rated_opponents": 653,
      "draws": 298,
      "entry_count": 8,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.7777777777777778,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:07Z",
      "losses": 211,
      "max_score": 85.25976029677969,
      "measured_at": null,
      "min_score": 9.848517519974454,
      "model_code_failure_rate": 0.1111111111111111,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": 0.8888888888888888,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1291,
      "rating_method": "sequential_elo",
      "raw_elo": 1702.2029115290434,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.1111111111111111,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 53.07968096604684,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 10,
      "wins": 782
    },
    {
      "admitted_plugins": 14,
      "attempted_matches": 899,
      "capped_draws": 92,
      "codegen_cost_usd": {
        "admitted_cell_count": 14,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 34
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 1.1371809,
        "mean": 0.8062893374999999,
        "median": 0.81347715,
        "min": 0.5199072,
        "recorded_attempt_count": 34,
        "run_count": 16,
        "total": 12.900629399999998
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": 54.22864747461619,
      "contributing_entrants": null,
      "distinct_rated_opponents": 418,
      "draws": 228,
      "entry_count": 14,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-02",
        "GB2-08"
      ],
      "infra_failures_excluded": 2,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:15Z",
      "losses": 267,
      "max_score": 82.92666095797935,
      "measured_at": null,
      "min_score": 23.77593276686914,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 2,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": 0.875,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 877,
      "rating_method": "sequential_elo",
      "raw_elo": 1610.2509525723651,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 14,
      "repaired_rate": 0.875,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 52.44822391600324,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 22,
      "wins": 382
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1466,
      "capped_draws": 22,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 1.38945,
        "mean": 0.93123625,
        "median": 0.95991,
        "min": 0.36561,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 7.449890000000001
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 51.47924961758838,
      "contributing_entrants": null,
      "distinct_rated_opponents": 650,
      "draws": 172,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-04",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 6,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:43:48Z",
      "losses": 230,
      "max_score": 51.47924961758838,
      "measured_at": null,
      "min_score": 51.47924961758838,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": 0.875,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1458,
      "rating_method": "sequential_elo",
      "raw_elo": 1828.7659874140832,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 49.78909371167585,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 8,
      "wins": 1056
    },
    {
      "admitted_plugins": 12,
      "attempted_matches": 2345,
      "capped_draws": 169,
      "codegen_cost_usd": {
        "admitted_cell_count": 12,
        "cell_count": 7,
        "cost_basis_counts": {
          "local_standard_estimate": 15
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 1.5164726249999998,
        "mean": 1.0182428035714286,
        "median": 0.9670455,
        "min": 0.8356068750000001,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 15
        },
        "run_count": 12,
        "token_priced_attempt_count": 15,
        "total": 12.214044
      },
      "codegen_evaluated_cells": 12,
      "conditional_success_score": 48.540299494544975,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1097,
      "draws": 616,
      "entry_count": 12,
      "failed_game_aliases": [],
      "first_pass_admitted": 10,
      "first_pass_rate": 0.8333333333333334,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-04"
      ],
      "infra_failures_excluded": 6,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:47Z",
      "losses": 606,
      "max_score": 80.51515814839607,
      "measured_at": null,
      "min_score": 12.309339079481154,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 2332,
      "rating_method": "sequential_elo",
      "raw_elo": 1598.956002424988,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.16666666666666666,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 48.540299494544975,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 13,
      "wins": 1110
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1309,
      "capped_draws": 142,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 10
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.454645,
        "mean": 0.28201406249999994,
        "median": 0.2695775,
        "min": 0.17401125,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 10
        },
        "run_count": 8,
        "token_priced_attempt_count": 10,
        "total": 2.2561125
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 44.65283798152629,
      "contributing_entrants": null,
      "distinct_rated_opponents": 656,
      "draws": 429,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:07Z",
      "losses": 305,
      "max_score": 82.16527921865402,
      "measured_at": null,
      "min_score": 14.429040501509393,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1293,
      "rating_method": "sequential_elo",
      "raw_elo": 1591.458962621409,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.25,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 44.65283798152629,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 10,
      "wins": 559
    },
    {
      "admitted_plugins": 15,
      "attempted_matches": 2806,
      "capped_draws": 389,
      "codegen_cost_usd": {
        "admitted_cell_count": 15,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 25
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 2,
        "max": 0.38568888333333334,
        "mean": 0.29249454166666666,
        "median": 0.2834616125,
        "min": 0.2238526,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 25
        },
        "run_count": 17,
        "token_priced_attempt_count": 25,
        "total": 5.07700365
      },
      "codegen_evaluated_cells": 17,
      "conditional_success_score": 45.64596858389405,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1292,
      "draws": 852,
      "entry_count": 15,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 10,
      "first_pass_rate": 0.5882352941176471,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07"
      ],
      "infra_failures_excluded": 9,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:39:21Z",
      "losses": 851,
      "max_score": 45.64596858389405,
      "measured_at": null,
      "min_score": 45.64596858389405,
      "model_code_failure_rate": 0.1176470588235294,
      "model_code_failures": 2,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": 0.8823529411764706,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 2775,
      "rating_method": "sequential_elo",
      "raw_elo": 1600.7989209775255,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.29411764705882354,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 44.23978538225884,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 22,
      "wins": 1072
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 967,
      "capped_draws": 106,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 18
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.6921525,
        "mean": 0.4890770625000001,
        "median": 0.483741,
        "min": 0.401826,
        "recorded_attempt_count": 18,
        "run_count": 8,
        "total": 3.912616499999999
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 45.710505493521154,
      "contributing_entrants": null,
      "distinct_rated_opponents": 475,
      "draws": 188,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:20Z",
      "losses": 297,
      "max_score": 87.65213640227793,
      "measured_at": null,
      "min_score": 5.549046286100777,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": 0.75,
      "provider": "Google",
      "provider_key": "google",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 932,
      "rating_method": "sequential_elo",
      "raw_elo": 1587.3523598129805,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.75,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 42.538418524284005,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 33,
      "wins": 447
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1260,
      "capped_draws": 67,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.3825,
        "mean": 0.27230375,
        "median": 0.25636749999999997,
        "min": 0.236235,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 2.17843
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 43.09225868476076,
      "contributing_entrants": null,
      "distinct_rated_opponents": 630,
      "draws": 350,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:22Z",
      "losses": 376,
      "max_score": 71.23639433663045,
      "measured_at": null,
      "min_score": 13.07288327052516,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": 0.875,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1259,
      "rating_method": "sequential_elo",
      "raw_elo": 1574.9480374312168,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 41.67746270276425,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 533
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1079,
      "capped_draws": 156,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 16
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.10909930499999998,
        "mean": 0.082847753875,
        "median": 0.084146916,
        "min": 0.03951746,
        "recorded_attempt_count": 16,
        "run_count": 8,
        "total": 0.662782031
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 44.70871398411597,
      "contributing_entrants": null,
      "distinct_rated_opponents": 539,
      "draws": 265,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 1,
      "first_pass_rate": 0.125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:09Z",
      "losses": 343,
      "max_score": 85.75114955106763,
      "measured_at": null,
      "min_score": 14.985507450020902,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": 0.75,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1078,
      "rating_method": "sequential_elo",
      "raw_elo": 1567.0799686218386,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 41.60614647782431,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 470
    },
    {
      "admitted_plugins": 13,
      "attempted_matches": 524,
      "capped_draws": 12,
      "codegen_cost_usd": {
        "admitted_cell_count": 13,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 26
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.184451,
        "mean": 0.0949985,
        "median": 0.08180399999999999,
        "min": 0.050682000000000005,
        "recorded_attempt_count": 26,
        "run_count": 16,
        "total": 1.5199759999999998
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": 43.36499484850174,
      "contributing_entrants": null,
      "distinct_rated_opponents": 262,
      "draws": 104,
      "entry_count": 13,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.4375,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:02Z",
      "losses": 139,
      "max_score": 59.14636993260279,
      "measured_at": null,
      "min_score": 11.001592922177778,
      "model_code_failure_rate": 0.1875,
      "model_code_failures": 3,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": 0.8125,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 524,
      "rating_method": "sequential_elo",
      "raw_elo": 1597.6214327952748,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 41.17135321287196,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 281
    },
    {
      "admitted_plugins": 10,
      "attempted_matches": 1587,
      "capped_draws": 154,
      "codegen_cost_usd": {
        "admitted_cell_count": 10,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 10
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.4567,
        "mean": 0.3660333333333334,
        "median": 0.355077,
        "min": 0.234622,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 10
        },
        "run_count": 10,
        "token_priced_attempt_count": 10,
        "total": 3.5443080000000005
      },
      "codegen_evaluated_cells": 10,
      "conditional_success_score": 40.57860418225832,
      "contributing_entrants": null,
      "distinct_rated_opponents": 673,
      "draws": 444,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 10,
      "first_pass_rate": 1.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 18,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:21Z",
      "losses": 435,
      "max_score": 64.79650419125173,
      "measured_at": null,
      "min_score": 0.0,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": 1.0,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1524,
      "rating_method": "sequential_elo",
      "raw_elo": 1568.0429525721834,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 40.57860418225832,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 53,
      "wins": 645
    },
    {
      "admitted_plugins": 19,
      "attempted_matches": 3362,
      "capped_draws": 418,
      "codegen_cost_usd": {
        "admitted_cell_count": 19,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 57
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 8,
        "max": 0.25293804045,
        "mean": 0.19953128278325,
        "median": 0.20662718916666667,
        "min": 0.13279406625,
        "recorded_attempt_count": 57,
        "run_count": 27,
        "total": 5.24138213703
      },
      "codegen_evaluated_cells": 27,
      "conditional_success_score": 38.86944403772714,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1586,
      "draws": 1101,
      "entry_count": 18,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-04",
        "GB2-06"
      ],
      "first_pass_admitted": 1,
      "first_pass_rate": 0.037037037037037035,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-03",
        "GB2-06"
      ],
      "infra_failures_excluded": 2,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:49:57Z",
      "losses": 929,
      "max_score": 62.93805242641539,
      "measured_at": null,
      "min_score": 9.910010721360797,
      "model_code_failure_rate": 0.2962962962962963,
      "model_code_failures": 8,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": 0.7037037037037037,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 3287,
      "rating_method": "sequential_elo",
      "raw_elo": 1575.5568138828494,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 18,
      "repaired_rate": 0.6666666666666666,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.600474625013355,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 49,
      "wins": 1257
    },
    {
      "admitted_plugins": 11,
      "attempted_matches": 2012,
      "capped_draws": 203,
      "codegen_cost_usd": {
        "admitted_cell_count": 11,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 26
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 4,
        "max": 0.015971877,
        "mean": 0.00979196948,
        "median": 0.009500477,
        "min": 0.003562685,
        "recorded_attempt_count": 26,
        "run_count": 15,
        "total": 0.14255357568000002
      },
      "codegen_evaluated_cells": 15,
      "conditional_success_score": 32.263197102394415,
      "contributing_entrants": null,
      "distinct_rated_opponents": 993,
      "draws": 510,
      "entry_count": 11,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-04"
      ],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.3333333333333333,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-08"
      ],
      "infra_failures_excluded": 1,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:54Z",
      "losses": 910,
      "max_score": 60.895193567762064,
      "measured_at": null,
      "min_score": 11.258167918522702,
      "model_code_failure_rate": 0.26666666666666666,
      "model_code_failures": 4,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": 0.7333333333333333,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 2006,
      "rating_method": "sequential_elo",
      "raw_elo": 1397.843684134,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.4,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 29.856078081711996,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 2,
      "wins": 586
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1286,
      "capped_draws": 79,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 13
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.0056201515,
        "mean": 0.0029446039375,
        "median": 0.001956843,
        "min": 0.001266762,
        "recorded_attempt_count": 13,
        "run_count": 9,
        "total": 0.029176983
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 29.722442719806377,
      "contributing_entrants": null,
      "distinct_rated_opponents": 634,
      "draws": 367,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-05"
      ],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.5555555555555556,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-19T20:23:50Z",
      "losses": 485,
      "max_score": 62.30809361425588,
      "measured_at": null,
      "min_score": 8.13506693682872,
      "model_code_failure_rate": 0.2222222222222222,
      "model_code_failures": 2,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": 0.7777777777777778,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1283,
      "rating_method": "sequential_elo",
      "raw_elo": 1486.7829836576914,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.2222222222222222,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 27.91247720515814,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 2,
      "wins": 431
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 872,
      "capped_draws": 131,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 18
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.01660212,
        "mean": 0.00855959175,
        "median": 0.005979701,
        "min": 0.00271606,
        "recorded_attempt_count": 18,
        "run_count": 8,
        "total": 0.06847673400000001
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 30.64330642131805,
      "contributing_entrants": null,
      "distinct_rated_opponents": 437,
      "draws": 396,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-05"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-13T11:17:15Z",
      "losses": 247,
      "max_score": 75.8950715607423,
      "measured_at": null,
      "min_score": 7.832227043501719,
      "model_code_failure_rate": 0.375,
      "model_code_failures": 3,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": 0.625,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 871,
      "rating_method": "sequential_elo",
      "raw_elo": 1490.89930790687,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 27.2461804322717,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 228
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 922,
      "capped_draws": 151,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 18
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.08193244999999999,
        "mean": 0.05495368125,
        "median": 0.056941025,
        "min": 0.02174195,
        "recorded_attempt_count": 18,
        "run_count": 8,
        "total": 0.43962945
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 28.64520773105748,
      "contributing_entrants": null,
      "distinct_rated_opponents": 461,
      "draws": 155,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:05:42Z",
      "losses": 529,
      "max_score": 49.13824204005262,
      "measured_at": null,
      "min_score": 6.917081520365711,
      "model_code_failure_rate": 0.375,
      "model_code_failures": 3,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": 0.625,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 922,
      "rating_method": "sequential_elo",
      "raw_elo": 1312.4609397125707,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 25.46959155221364,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 238
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 876,
      "capped_draws": 208,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 17
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.0597184625,
        "mean": 0.02441709805,
        "median": 0.022888967499999996,
        "min": 0.0040410944,
        "recorded_attempt_count": 17,
        "run_count": 8,
        "total": 0.1953367844
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 24.521755430704992,
      "contributing_entrants": null,
      "distinct_rated_opponents": 439,
      "draws": 487,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-05",
        "GB2-07"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:37:33Z",
      "losses": 264,
      "max_score": 51.50438313656388,
      "measured_at": null,
      "min_score": 11.490618140835323,
      "model_code_failure_rate": 0.375,
      "model_code_failures": 3,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": 0.625,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 876,
      "rating_method": "sequential_elo",
      "raw_elo": 1415.4539371566143,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 21.80326639021641,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 125
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 891,
      "capped_draws": 127,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 18
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.0422995,
        "mean": 0.0264053125,
        "median": 0.022682375,
        "min": 0.01946775,
        "recorded_attempt_count": 18,
        "run_count": 8,
        "total": 0.2112425
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 23.913526464397528,
      "contributing_entrants": null,
      "distinct_rated_opponents": 445,
      "draws": 266,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:30Z",
      "losses": 451,
      "max_score": 50.07010386016034,
      "measured_at": null,
      "min_score": 6.198198890963618,
      "model_code_failure_rate": 0.375,
      "model_code_failures": 3,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": 0.625,
      "provider": "Google",
      "provider_key": "google",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 887,
      "rating_method": "sequential_elo",
      "raw_elo": 1314.551923630182,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 21.2624658665295,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 4,
      "wins": 170
    },
    {
      "admitted_plugins": 9,
      "attempted_matches": 1785,
      "capped_draws": 316,
      "codegen_cost_usd": {
        "admitted_cell_count": 9,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 33
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 7,
        "max": 0.039679950000000005,
        "mean": 0.029276065625,
        "median": 0.0306386,
        "min": 0.01864685,
        "recorded_attempt_count": 33,
        "run_count": 16,
        "total": 0.46841705
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": 21.843112445435427,
      "contributing_entrants": null,
      "distinct_rated_opponents": 832,
      "draws": 776,
      "entry_count": 9,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-04",
        "GB2-06",
        "GB2-07"
      ],
      "first_pass_admitted": 2,
      "first_pass_rate": 0.125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-19T20:28:27Z",
      "losses": 691,
      "max_score": 36.33021415855389,
      "measured_at": null,
      "min_score": 10.307786108744647,
      "model_code_failure_rate": 0.4375,
      "model_code_failures": 7,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": 0.5625,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1757,
      "rating_method": "sequential_elo",
      "raw_elo": 1377.446118812076,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 7,
      "repaired_rate": 0.4375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 18.91669027546711,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 13,
      "wins": 290
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 856,
      "capped_draws": 8,
      "codegen_cost_usd": {
        "admitted_cell_count": 4,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 4,
        "max": 0.055222,
        "mean": 0.046879,
        "median": 0.046959,
        "min": 0.039474,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 0.375032
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 64.20944368278158,
      "contributing_entrants": null,
      "distinct_rated_opponents": 383,
      "draws": 19,
      "entry_count": 4,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.5,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 24,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:18Z",
      "losses": 332,
      "max_score": 74.07987912466021,
      "measured_at": null,
      "min_score": 37.904090563653,
      "model_code_failure_rate": 0.5,
      "model_code_failures": 4,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": 0.5,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": null,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "provisional",
      "rated_matches": 850,
      "rating_method": "sequential_elo",
      "raw_elo": 1588.2823551530091,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 53.9934910182863,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 6,
      "wins": 499
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 727,
      "capped_draws": 32,
      "codegen_cost_usd": {
        "admitted_cell_count": 4,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 13
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 5,
        "max": 0.33185775,
        "mean": 0.18266821875,
        "median": 0.16804049999999998,
        "min": 0.08751,
        "recorded_attempt_count": 13,
        "run_count": 9,
        "total": 1.7932035
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 18.35955162608331,
      "contributing_entrants": null,
      "distinct_rated_opponents": 364,
      "draws": 32,
      "entry_count": 4,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-08"
      ],
      "first_pass_admitted": 2,
      "first_pass_rate": 0.2222222222222222,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-13T11:15:27Z",
      "losses": 577,
      "max_score": 23.80090952267913,
      "measured_at": null,
      "min_score": 10.384314501114984,
      "model_code_failure_rate": 0.5555555555555556,
      "model_code_failures": 5,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": 0.4444444444444444,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": null,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "provisional",
      "rated_matches": 727,
      "rating_method": "sequential_elo",
      "raw_elo": 1128.2233049327087,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.2222222222222222,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 14.990511130063094,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 118
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 201,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 0,
      "max_score": null,
      "measured_at": "2026-07-19T18:25:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 201,
      "rating_method": "sequential_elo",
      "raw_elo": 2217.851148747644,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 100.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 201
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 19,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1895.3679746282792,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 74.0960959782584,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 157
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 42,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 42,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T20:13:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1881.636733207583,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 73.13624142561254,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 145
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 24,
      "max_score": null,
      "measured_at": "2026-07-19T17:08:13Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1871.3032618400064,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 72.31110877690001,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 174
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 34,
      "max_score": null,
      "measured_at": "2026-07-19T19:40:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1863.773097164302,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 71.73223969080041,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 168
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 28,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 28,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 36,
      "max_score": null,
      "measured_at": "2026-07-19T19:31:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1762.925387581977,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 63.67950167428481,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 138
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 21,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 21,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 36,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1747.742032182051,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.30809361425588,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 119
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 22,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 22,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-19T20:13:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 201,
      "rating_method": "sequential_elo",
      "raw_elo": 1732.0159014279725,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 61.205822303176355,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 136
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 80,
      "capped_draws": 11,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 11,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-18T13:38:53Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1729.047179584418,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.22785256603646,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 63
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 178,
      "capped_draws": 29,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 29,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 73,
      "max_score": null,
      "measured_at": "2026-07-18T12:50:58Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 178,
      "rating_method": "sequential_elo",
      "raw_elo": 1557.5991820583583,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 47.13854557235755,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 76
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 408,
      "capped_draws": 221,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 183,
      "draws": 221,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 105,
      "max_score": null,
      "measured_at": "2026-07-19T20:19:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 404,
      "rating_method": "sequential_elo",
      "raw_elo": 1538.7722743262907,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 45.780747588218205,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 78
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 27,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 60,
      "draws": 27,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 51,
      "max_score": null,
      "measured_at": "2026-07-19T20:19:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 127,
      "rating_method": "sequential_elo",
      "raw_elo": 1555.386633078287,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 45.37208589267169,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 18.508683647302114,
      "unilateral_faults": 1,
      "wins": 49
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 15,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 15,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 80,
      "max_score": null,
      "measured_at": "2026-07-19T19:43:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 1501.9828344124348,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.86487874768688,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 111
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 178,
      "capped_draws": 32,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 32,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 70,
      "max_score": null,
      "measured_at": "2026-07-18T07:07:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 178,
      "rating_method": "sequential_elo",
      "raw_elo": 1443.5446665239044,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 38.03123776136748,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 76
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 38,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 38,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 80,
      "max_score": null,
      "measured_at": "2026-07-13T09:49:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1431.1821881901408,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.03063812445185,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 58
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 365,
      "capped_draws": 74,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 183,
      "draws": 74,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 182,
      "max_score": null,
      "measured_at": "2026-07-18T13:38:53Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 365,
      "rating_method": "sequential_elo",
      "raw_elo": 1393.7558331773266,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.083919098748204,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 109
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 50,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 50,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 120,
      "max_score": null,
      "measured_at": "2026-07-19T20:13:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 203,
      "rating_method": "sequential_elo",
      "raw_elo": 1340.0696408267977,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 29.91976859210062,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 33
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 67,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 67,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 90,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1326.3495764650936,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 28.659703759107938,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 19
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 379,
      "capped_draws": 105,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 179,
      "draws": 105,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 229,
      "max_score": null,
      "measured_at": "2026-07-19T19:43:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 379,
      "rating_method": "sequential_elo",
      "raw_elo": 1303.5716447340765,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 26.923756518498394,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 45
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 75,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 75,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 100,
      "max_score": null,
      "measured_at": "2026-07-13T10:53:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1212.999707894921,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.615415585825644,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 2
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 32,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 32,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 131,
      "max_score": null,
      "measured_at": "2026-07-13T11:04:49Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1097.4794524632837,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.384314501114984,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 13
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 13,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 147,
      "max_score": null,
      "measured_at": "2026-07-13T10:56:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1054.0578852920023,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 6.917081520365711,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 16
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 83,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T20:22:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1991.6326427339416,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 100.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 164
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 64,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 3,
      "max_score": null,
      "measured_at": "2026-07-19T20:15:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1761.809414029661,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.54693366398842,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 119
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 77,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 3,
      "max_score": null,
      "measured_at": "2026-07-19T19:50:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1741.2859310206684,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 57.03469150252695,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 107
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 119,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-19T19:25:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1611.4657516747402,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.73354815308172,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 64
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 143,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T20:20:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1555.4872213111505,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.090411884130972,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 31
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 83,
      "draws": 132,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 12,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 182,
      "rating_method": "sequential_elo",
      "raw_elo": 1545.289767505208,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.297513470769697,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 4,
      "wins": 38
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 62,
      "draws": 83,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 15,
      "max_score": null,
      "measured_at": "2026-07-19T20:29:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 124,
      "rating_method": "sequential_elo",
      "raw_elo": 1549.741780384856,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.48463931663812,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 19.244078834871203,
      "unilateral_faults": 4,
      "wins": 26
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 349,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 165,
      "draws": 301,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 28,
      "max_score": null,
      "measured_at": "2026-07-19T20:20:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 347,
      "rating_method": "sequential_elo",
      "raw_elo": 1496.5288406023028,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.791801001532535,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 18
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 172,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 144,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-18T13:07:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 172,
      "rating_method": "sequential_elo",
      "raw_elo": 1494.5068949772526,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.429040501509393,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 11
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 151,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-13T03:20:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1487.3409069796846,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 13.07288327052516,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 5
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 350,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 165,
      "draws": 303,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 32,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 348,
      "rating_method": "sequential_elo",
      "raw_elo": 1482.030450946625,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.309339079481154,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 13
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 162,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 147,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-13T02:43:13Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1478.3206312567197,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.490618140835323,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 7
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 40,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 34,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 2,
      "max_score": null,
      "measured_at": "2026-07-18T13:31:39Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 40,
      "rating_method": "sequential_elo",
      "raw_elo": 1508.952275545118,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.4034471997158,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 4
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 350,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 165,
      "draws": 306,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-19T20:29:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 348,
      "rating_method": "sequential_elo",
      "raw_elo": 1475.9113373881587,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.258167918522702,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 12
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 350,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 165,
      "draws": 312,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 32,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 348,
      "rating_method": "sequential_elo",
      "raw_elo": 1470.3789422522132,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.307786108744647,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 4
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 520,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 250,
      "draws": 439,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 53,
      "max_score": null,
      "measured_at": "2026-07-19T20:19:16Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 519,
      "rating_method": "sequential_elo",
      "raw_elo": 1468.1418423162531,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 9.910010721360797,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 27
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 162,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 145,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-13T03:20:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1468.7615788904866,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 9.848517519974454,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 136,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 23,
      "max_score": null,
      "measured_at": "2026-07-13T03:35:49Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1458.6915160586964,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 8.13506693682872,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 144,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 19,
      "max_score": null,
      "measured_at": "2026-07-13T03:45:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1456.9286141012149,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.832227043501719,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 51,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 25,
      "draws": 6,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 3,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:08Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 37,
      "rating_method": "sequential_elo",
      "raw_elo": 1726.1253813872777,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 100.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 37.360710565381275,
      "unilateral_faults": 8,
      "wins": 28
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 57,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 24,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 42,
      "rating_method": "sequential_elo",
      "raw_elo": 1656.8215338050852,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 83.07600646134895,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 32.49928516630233,
      "unilateral_faults": 6,
      "wins": 31
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 60,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 18,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 0,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 26,
      "rating_method": "sequential_elo",
      "raw_elo": 1592.3879631862185,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.27137257662854,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 52.475044864937615,
      "unilateral_faults": 34,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 78,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 36,
      "draws": 25,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 61,
      "rating_method": "sequential_elo",
      "raw_elo": 1559.9612376674422,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.65522539740799,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 19.750063500095248,
      "unilateral_faults": 5,
      "wins": 22
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 62,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 23,
      "draws": 21,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 2,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:08Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 28,
      "rating_method": "sequential_elo",
      "raw_elo": 1548.056257492551,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.50097274542412,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 49.09766908852594,
      "unilateral_faults": 33,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 53,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 30,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 39,
      "rating_method": "sequential_elo",
      "raw_elo": 1536.3665196595405,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.33095440558088,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 35.30694150420948,
      "unilateral_faults": 10,
      "wins": 14
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 53,
      "capped_draws": 2,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 27,
      "draws": 20,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 37,
      "rating_method": "sequential_elo",
      "raw_elo": 1472.9634286487662,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.502324757678714,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 37.360710565381275,
      "unilateral_faults": 10,
      "wins": 6
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 77,
      "capped_draws": 2,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 26,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:08Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 34,
      "rating_method": "sequential_elo",
      "raw_elo": 1447.0819507113183,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.101618034124503,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 40.76542547285167,
      "unilateral_faults": 37,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 57,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 5,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 21,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 37,
      "rating_method": "sequential_elo",
      "raw_elo": 1379.346598885276,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.281920993460163,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 37.360710565381275,
      "unilateral_faults": 11,
      "wins": 11
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 64,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 25,
      "draws": 5,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 20,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 35,
      "rating_method": "sequential_elo",
      "raw_elo": 1379.961099285522,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 9.980876737333936,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 39.583333333333336,
      "unilateral_faults": 10,
      "wins": 10
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 83,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 25,
      "draws": 4,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 22,
      "max_score": null,
      "measured_at": "2026-07-19T23:13:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 31,
      "rating_method": "sequential_elo",
      "raw_elo": 1345.4269744064488,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 0.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 44.63834764831843,
      "unilateral_faults": 42,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 23,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 3,
      "max_score": null,
      "measured_at": "2026-07-19T20:25:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1910.0429790696871,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 100.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 160
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 84,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-19T20:11:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1710.7933885474688,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.7273441167311,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 92
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 83,
      "draws": 85,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-19T20:20:09Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1706.679265118354,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 61.943045486748765,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 92
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 126,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-19T20:25:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1573.5715975142334,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.04323878185268,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 45
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 162,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 92,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-13T06:44:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1558.506800704954,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 33.85082376288825,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 54
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 162,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 122,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 22,
      "max_score": null,
      "measured_at": "2026-07-13T05:17:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1513.9665371375488,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.51889237699729,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 18
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 106,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 33,
      "max_score": null,
      "measured_at": "2026-07-13T05:27:45Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1512.9440218097536,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.345511091461717,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 24
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 84,
      "draws": 89,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 25,
      "max_score": null,
      "measured_at": "2026-07-19T20:21:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1507.7468988933458,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 24.744420702191245,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 72
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 83,
      "draws": 129,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 32,
      "max_score": null,
      "measured_at": "2026-07-19T20:09:46Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1507.673579789397,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 24.71601576668925,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 24
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 60,
      "draws": 45,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 32,
      "max_score": null,
      "measured_at": "2026-07-19T20:20:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 124,
      "rating_method": "sequential_elo",
      "raw_elo": 1523.7139890557742,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.77593276686914,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 19.244078834871203,
      "unilateral_faults": 4,
      "wins": 47
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 174,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 87,
      "draws": 110,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 29,
      "max_score": null,
      "measured_at": "2026-07-18T13:31:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 174,
      "rating_method": "sequential_elo",
      "raw_elo": 1495.1848218252635,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.209998741054164,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 35
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 152,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 18,
      "max_score": null,
      "measured_at": "2026-07-18T13:32:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1490.1018224099944,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.291194047877852,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 124,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 23,
      "max_score": null,
      "measured_at": "2026-07-13T04:51:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1484.8345438109034,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.10493995228072,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 17
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 83,
      "draws": 116,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-19T20:09:46Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1478.4234125991477,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.24432865630252,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 21
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 139,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 15,
      "max_score": null,
      "measured_at": "2026-07-13T04:36:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1475.3273603281289,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.32647718383133,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 120,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 29,
      "max_score": null,
      "measured_at": "2026-07-13T05:12:08Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1462.0402675941834,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.823192329999186,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 14
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 348,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 164,
      "draws": 226,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 72,
      "max_score": null,
      "measured_at": "2026-07-19T20:20:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 346,
      "rating_method": "sequential_elo",
      "raw_elo": 1457.4590105074017,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.128057056937948,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 48
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 384,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 170,
      "draws": 151,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 152,
      "max_score": null,
      "measured_at": "2026-07-19T20:21:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 382,
      "rating_method": "sequential_elo",
      "raw_elo": 1450.4644820831422,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.099220120745768,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 79
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 116,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 29,
      "max_score": null,
      "measured_at": "2026-07-13T06:16:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1444.813470760353,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.600658879990688,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 18
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 80,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 55,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-18T13:27:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1467.5000467400714,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.001592922177778,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 55,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 65,
      "max_score": null,
      "measured_at": "2026-07-13T06:10:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1434.8720577315091,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.740966893440644,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 43
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 83,
      "draws": 141,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 37,
      "max_score": null,
      "measured_at": "2026-07-19T20:21:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1413.940355124856,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.166986972807464,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 214,
      "rating_method": "sequential_elo",
      "raw_elo": 2189.4172617110444,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 100.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 208
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 192,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 2101.5599256868545,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 93.07296360431104,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 187
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 192,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 2007.0964504243177,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 85.75114955106763,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 176
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 192,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 23,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 1960.8328098869608,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 82.16527921865402,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 169
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-19T20:24:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 215,
      "rating_method": "sequential_elo",
      "raw_elo": 1933.679334910949,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.18278864235208,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 185
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 213,
      "rating_method": "sequential_elo",
      "raw_elo": 1933.1657077330144,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.13314292967871,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 183
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 408,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 194,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 73,
      "max_score": null,
      "measured_at": "2026-07-19T20:24:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 406,
      "rating_method": "sequential_elo",
      "raw_elo": 1930.4565549318795,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 79.8694620999203,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 333
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 35,
      "max_score": null,
      "measured_at": "2026-07-19T18:36:49Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 216,
      "rating_method": "sequential_elo",
      "raw_elo": 1924.5872917457057,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 79.48293548632148,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 181
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 68,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:50Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 214,
      "rating_method": "sequential_elo",
      "raw_elo": 1738.7297749558174,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 65.06745104418093,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 146
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 217,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 77,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:50Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 216,
      "rating_method": "sequential_elo",
      "raw_elo": 1735.1081095036748,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 64.79650419125173,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 139
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 58,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-19T20:26:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 126,
      "rating_method": "sequential_elo",
      "raw_elo": 1728.4092662022958,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.50648111946364,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 18.75,
      "unilateral_faults": 2,
      "wins": 81
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 194,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 82,
      "max_score": null,
      "measured_at": "2026-07-18T07:13:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 194,
      "rating_method": "sequential_elo",
      "raw_elo": 1636.486916757509,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 57.03687454599186,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 112
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 194,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 65,
      "max_score": null,
      "measured_at": "2026-07-18T07:08:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 194,
      "rating_method": "sequential_elo",
      "raw_elo": 1630.3002526236642,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 56.557349478001406,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 129
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 420,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 198,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 165,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 419,
      "rating_method": "sequential_elo",
      "raw_elo": 1627.412309265274,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 56.41124513301693,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 254
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 81,
      "max_score": null,
      "measured_at": "2026-07-19T20:15:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 215,
      "rating_method": "sequential_elo",
      "raw_elo": 1611.284714442574,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.19415196267263,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 134
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 196,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 81,
      "max_score": null,
      "measured_at": "2026-07-18T07:45:17Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 196,
      "rating_method": "sequential_elo",
      "raw_elo": 1589.2056385127698,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 53.383426593216655,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 115
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 201,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 101,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 85,
      "max_score": null,
      "measured_at": "2026-07-18T12:52:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 201,
      "rating_method": "sequential_elo",
      "raw_elo": 1564.2836972542352,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.47924961758838,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 116
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 615,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 298,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 254,
      "max_score": null,
      "measured_at": "2026-07-19T20:19:26Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 609,
      "rating_method": "sequential_elo",
      "raw_elo": 1562.8811661980608,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.37851771163366,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 355
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 100,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 107,
      "max_score": null,
      "measured_at": "2026-07-18T13:05:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 200,
      "rating_method": "sequential_elo",
      "raw_elo": 1534.1507619303593,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 49.13824204005262,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 93
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 80,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 36,
      "max_score": null,
      "measured_at": "2026-07-18T15:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1555.026515631915,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.10117101320647,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 44
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 428,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 202,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 267,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 421,
      "rating_method": "sequential_elo",
      "raw_elo": 1368.2761811372434,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.33021415855389,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 154
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 384,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 192,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 231,
      "max_score": null,
      "measured_at": "2026-07-13T10:59:39Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 384,
      "rating_method": "sequential_elo",
      "raw_elo": 1366.2448487733957,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.07908053211064,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 153
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 192,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 147,
      "max_score": null,
      "measured_at": "2026-07-13T10:00:53Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 1207.836195537406,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.80090952267913,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 45
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 192,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 164,
      "max_score": null,
      "measured_at": "2026-07-13T10:58:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 980.732200786807,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 6.198198890963618,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 28
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 201,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 200,
      "rating_method": "sequential_elo",
      "raw_elo": 2113.2529956970648,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 100.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 180
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 7,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 22,
      "max_score": null,
      "measured_at": "2026-07-19T20:26:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 199,
      "rating_method": "sequential_elo",
      "raw_elo": 2046.7163075917915,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 94.51620294358332,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 170
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 24,
      "max_score": null,
      "measured_at": "2026-07-19T20:21:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 2027.4878813095452,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 92.92727390319158,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 172
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:24Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 2015.6860921297173,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 91.81479306316908,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 153
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 7,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1965.1257540745448,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 87.65213640227793,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 155
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 25,
      "max_score": null,
      "measured_at": "2026-07-19T20:11:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 199,
      "rating_method": "sequential_elo",
      "raw_elo": 1962.620486155873,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 87.59255399821146,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 162
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 27,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:50Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 203,
      "rating_method": "sequential_elo",
      "raw_elo": 1941.2653628513624,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 85.85731954540609,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 168
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 19,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:13Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1936.06754569874,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 85.25976029677969,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 143
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 6,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-19T19:49:17Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 200,
      "rating_method": "sequential_elo",
      "raw_elo": 1921.6515965115368,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 84.22536574645959,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 164
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 60,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 124,
      "rating_method": "sequential_elo",
      "raw_elo": 1928.063507325252,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 82.92666095797935,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 19.250031750047626,
      "unilateral_faults": 4,
      "wins": 100
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 5,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 42,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1822.3224384637483,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 75.8950715607423,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 129
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 47,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1770.4153570626615,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 71.76237080818703,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 150
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 37,
      "max_score": null,
      "measured_at": "2026-07-13T11:01:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1758.9621117388845,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 70.67858574947873,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 131
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 17,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 84,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 207,
      "rating_method": "sequential_elo",
      "raw_elo": 1592.5756149645515,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 57.171804844224184,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 106
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 376,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 178,
      "draws": 15,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 147,
      "max_score": null,
      "measured_at": "2026-07-19T20:26:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 373,
      "rating_method": "sequential_elo",
      "raw_elo": 1567.9514463506266,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.02005996423915,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 211
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 82,
      "max_score": null,
      "measured_at": "2026-07-18T13:37:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1525.0221840943957,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.50438313656388,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 97
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 80,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 3,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 35,
      "max_score": null,
      "measured_at": "2026-07-18T13:59:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1557.562383696878,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.42944139758407,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 42
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 88,
      "max_score": null,
      "measured_at": "2026-07-18T13:19:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1492.9147765362331,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.86096501324975,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 89
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 11,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 109,
      "max_score": null,
      "measured_at": "2026-07-19T20:21:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 200,
      "rating_method": "sequential_elo",
      "raw_elo": 1359.0254845881047,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.904090563653,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 80
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 3,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 108,
      "max_score": null,
      "measured_at": "2026-07-13T11:02:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1316.3454906374743,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.23774898057906,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 65
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 4,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 110,
      "max_score": null,
      "measured_at": "2026-07-13T10:53:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1315.3525016186966,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.1559957227449,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 62
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 7,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 127,
      "max_score": null,
      "measured_at": "2026-07-19T20:24:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1284.9657247905943,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.818186314401817,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 68
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 130,
      "max_score": null,
      "measured_at": "2026-07-13T10:57:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1243.9983849389728,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 28.281377385447776,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 46
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 143,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1175.6040568173992,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.650439846783147,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 33
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 145,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:58Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1146.417131795777,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.247466438662197,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 31
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 352,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 176,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 286,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:39Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 352,
      "rating_method": "sequential_elo",
      "raw_elo": 1131.260629418715,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.999624394828466,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 65
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 213,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 212,
      "rating_method": "sequential_elo",
      "raw_elo": 2213.60163755207,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 100.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 207
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 214,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 213,
      "rating_method": "sequential_elo",
      "raw_elo": 2191.0952211856857,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 98.42557971783368,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 208
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 276,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 102,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 50,
      "max_score": null,
      "measured_at": "2026-07-20T08:43:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 275,
      "rating_method": "sequential_elo",
      "raw_elo": 1961.8313657037445,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 82.5709165265804,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 225
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 398,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 184,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 56,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 395,
      "rating_method": "sequential_elo",
      "raw_elo": 1936.877498008331,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.51515814839607,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 339
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 184,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 28,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1924.4565650704055,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 79.57489937279772,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 156
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 207,
      "rating_method": "sequential_elo",
      "raw_elo": 1866.9421353328428,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 75.65743305436055,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 164
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 36,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1829.7985900997423,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 72.93980546270726,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 149
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 41,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 207,
      "rating_method": "sequential_elo",
      "raw_elo": 1818.0107769229128,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 72.22469640238779,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 166
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 185,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 52,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1744.5797328806148,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 66.95579753782103,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 132
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 183,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 54,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:20Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 183,
      "rating_method": "sequential_elo",
      "raw_elo": 1723.0027444301477,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 65.4364840406668,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 129
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 595,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 285,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 205,
      "max_score": null,
      "measured_at": "2026-07-20T08:49:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 585,
      "rating_method": "sequential_elo",
      "raw_elo": 1686.5017799498585,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.93805242641539,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 380
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 67,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1658.305456631933,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.914363604458195,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 119
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 375,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 184,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 132,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 371,
      "rating_method": "sequential_elo",
      "raw_elo": 1658.0752540736505,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.895193567762064,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 239
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 183,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 65,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:09Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 183,
      "rating_method": "sequential_elo",
      "raw_elo": 1656.8247006220377,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.79382125920891,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 118
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 84,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 42,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 26,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 84,
      "rating_method": "sequential_elo",
      "raw_elo": 1664.8524239181024,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 59.14636993260279,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 32.49928516630233,
      "unilateral_faults": 0,
      "wins": 58
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 131,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 58,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 57,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 127,
      "rating_method": "sequential_elo",
      "raw_elo": 1616.8225752179756,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 56.562076636097146,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 18.508683647302114,
      "unilateral_faults": 4,
      "wins": 70
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 289,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 101,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 142,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 287,
      "rating_method": "sequential_elo",
      "raw_elo": 1560.1947024692745,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 54.43001195791607,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 145
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 419,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 186,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 208,
      "max_score": null,
      "measured_at": "2026-07-20T08:39:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 417,
      "rating_method": "sequential_elo",
      "raw_elo": 1563.500079681622,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 54.37649620014344,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 209
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 173,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 89,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 169,
      "rating_method": "sequential_elo",
      "raw_elo": 1505.155510529379,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 50.07010386016034,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 4,
      "wins": 80
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 225,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 118,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 224,
      "rating_method": "sequential_elo",
      "raw_elo": 1482.9798355750318,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.795842422159346,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 106
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 184,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 120,
      "max_score": null,
      "measured_at": "2026-07-18T13:03:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1365.4718607095751,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 40.359816348203495,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 64
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 182,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 111,
      "max_score": null,
      "measured_at": "2026-07-13T10:08:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 182,
      "rating_method": "sequential_elo",
      "raw_elo": 1362.3074404228162,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 40.12657589285319,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 71
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 182,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 122,
      "max_score": null,
      "measured_at": "2026-07-13T10:52:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 182,
      "rating_method": "sequential_elo",
      "raw_elo": 1259.866347632146,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.9399106271018,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 60
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 183,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 154,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 183,
      "rating_method": "sequential_elo",
      "raw_elo": 1061.1604399343687,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.005516041876938,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 29
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 2126.6923718578123,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 98.290012518652,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 184
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 213,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 15,
      "max_score": null,
      "measured_at": "2026-07-19T20:20:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 211,
      "rating_method": "sequential_elo",
      "raw_elo": 2043.1402636259547,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 89.47819347812705,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 196
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 213,
      "capped_draws": 8,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 24,
      "max_score": null,
      "measured_at": "2026-07-19T20:26:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 211,
      "rating_method": "sequential_elo",
      "raw_elo": 1899.8393860063263,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 74.07987912466021,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 179
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 27,
      "max_score": null,
      "measured_at": "2026-07-13T11:01:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1874.9240848051988,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 71.23639433663045,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 162
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 219,
      "capped_draws": 69,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 100,
      "draws": 69,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-19T20:26:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 216,
      "rating_method": "sequential_elo",
      "raw_elo": 1823.0540236330833,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 65.86315907036878,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 131
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 9,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 29,
      "max_score": null,
      "measured_at": "2026-07-13T10:32:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1794.52112484674,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.58850158383014,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 150
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 218,
      "capped_draws": 40,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 100,
      "draws": 40,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 52,
      "max_score": null,
      "measured_at": "2026-07-19T19:54:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 218,
      "rating_method": "sequential_elo",
      "raw_elo": 1670.0153149148105,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 49.43182542421367,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 126
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 80,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-18T15:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1700.4092044504214,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.889597458780386,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 62
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 610,
      "capped_draws": 283,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 286,
      "draws": 283,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 118,
      "max_score": null,
      "measured_at": "2026-07-19T20:26:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 608,
      "rating_method": "sequential_elo",
      "raw_elo": 1635.8068593101143,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 45.64596858389405,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 207
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 196,
      "capped_draws": 32,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 32,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-18T12:51:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 196,
      "rating_method": "sequential_elo",
      "raw_elo": 1607.049252758707,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.50795753707946,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 121
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 409,
      "capped_draws": 154,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 195,
      "draws": 154,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 94,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 406,
      "rating_method": "sequential_elo",
      "raw_elo": 1568.763898202228,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 38.44569514759158,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 158
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 591,
      "capped_draws": 195,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 286,
      "draws": 195,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 199,
      "max_score": null,
      "measured_at": "2026-07-19T20:19:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 587,
      "rating_method": "sequential_elo",
      "raw_elo": 1502.3511242022505,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.251195291498743,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 193
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 194,
      "capped_draws": 58,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 58,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 77,
      "max_score": null,
      "measured_at": "2026-07-18T12:53:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 194,
      "rating_method": "sequential_elo",
      "raw_elo": 1431.1398224703053,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.590045524889156,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 59
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 214,
      "capped_draws": 126,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 126,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 49,
      "max_score": null,
      "measured_at": "2026-07-19T20:28:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 211,
      "rating_method": "sequential_elo",
      "raw_elo": 1419.0540128261568,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.41736012440615,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 36
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 138,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 138,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-18T07:37:39Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1398.8772020896483,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.07480874502237,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 7
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 109,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 109,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 55,
      "max_score": null,
      "measured_at": "2026-07-13T11:02:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1394.3186977418932,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.56830265888095,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 22
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 405,
      "capped_draws": 266,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 194,
      "draws": 266,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 108,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 399,
      "rating_method": "sequential_elo",
      "raw_elo": 1387.0458687407518,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.891337069007747,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 3,
      "wins": 25
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 131,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 131,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 46,
      "max_score": null,
      "measured_at": "2026-07-13T11:13:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1368.1245757862173,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.761998729502363,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 127,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 127,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 53,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1367.5461615676222,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.691475001584593,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 129,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 129,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 49,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:20Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1361.814201972755,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.075550541321224,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 127,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 127,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-13T11:13:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1351.6699661790897,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.985507450020902,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 11
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 65,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 60,
      "draws": 65,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 51,
      "max_score": null,
      "measured_at": "2026-07-19T20:23:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K3",
      "model_slug": "moonshotai-kimi-k3",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 125,
      "rating_method": "sequential_elo",
      "raw_elo": 1369.6189167421148,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k3",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.66338227359166,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 18.9970394174356,
      "unilateral_faults": 3,
      "wins": 9
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 133,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 133,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 52,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:45Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1346.9606253994864,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.479467913302816,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 1
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 212,
      "capped_draws": 29,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 29,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 137,
      "max_score": null,
      "measured_at": "2026-07-19T20:26:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 210,
      "rating_method": "sequential_elo",
      "raw_elo": 1281.243785441023,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.602053478392984,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 2,
      "wins": 44
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 105,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 105,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 81,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1263.85170663636,
      "reasoning_level": "XHigh",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "highest",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/highest"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "highest",
        "reasoning_bucket_label": "Highest reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "highest"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 5.549046286100777,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/highest",
      "track_name": "XHigh variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1355,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.6981,
        "mean": 0.5046499999999999,
        "median": 0.511355,
        "min": 0.33027,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 4.0372
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 71.77619468353247,
      "contributing_entrants": null,
      "distinct_rated_opponents": 628,
      "draws": 181,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:02Z",
      "losses": 137,
      "max_score": 85.85876169674393,
      "measured_at": null,
      "min_score": 44.00470514710234,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": 0.875,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1349,
      "rating_method": "sequential_elo",
      "raw_elo": 1862.005018618128,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 69.41965374228975,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 1031
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1462,
      "capped_draws": 18,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 10
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.6689649999999999,
        "mean": 0.46799875,
        "median": 0.4352825,
        "min": 0.366615,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 10
        },
        "run_count": 9,
        "token_priced_attempt_count": 10,
        "total": 4.1199200000000005
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 64.64087807182851,
      "contributing_entrants": null,
      "distinct_rated_opponents": 639,
      "draws": 247,
      "entry_count": 8,
      "failed_game_aliases": [
        "GB2-06"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.7777777777777778,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:39:18Z",
      "losses": 177,
      "max_score": 90.58621353721448,
      "measured_at": null,
      "min_score": 25.456448781739248,
      "model_code_failure_rate": 0.1111111111111111,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": 0.8888888888888888,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1259,
      "rating_method": "sequential_elo",
      "raw_elo": 1764.368616883458,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.1111111111111111,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 62.76522883961819,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 195,
      "wins": 835
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1414,
      "capped_draws": 126,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 19
        },
        "evidence_basis": "Provider-reported",
        "max": 0.5145615,
        "mean": 0.34875300000000004,
        "median": 0.335163,
        "min": 0.254883,
        "recorded_attempt_count": 19,
        "run_count": 8,
        "total": 2.790024
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 52.16906430603814,
      "contributing_entrants": null,
      "distinct_rated_opponents": 658,
      "draws": 415,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:33Z",
      "losses": 296,
      "max_score": 91.79025106882118,
      "measured_at": null,
      "min_score": 15.64390388899189,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": 1.0,
      "provider": "Google",
      "provider_key": "google",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1390,
      "rating_method": "sequential_elo",
      "raw_elo": 1667.2044190421009,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 8,
      "repaired_rate": 1.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 52.16906430603814,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 13,
      "wins": 679
    },
    {
      "admitted_plugins": 15,
      "attempted_matches": 834,
      "capped_draws": 20,
      "codegen_cost_usd": {
        "admitted_cell_count": 15,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 30
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.141566,
        "mean": 0.10599,
        "median": 0.105677,
        "min": 0.046481,
        "recorded_attempt_count": 30,
        "run_count": 16,
        "total": 1.69584
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": 51.00775136236891,
      "contributing_entrants": null,
      "distinct_rated_opponents": 302,
      "draws": 141,
      "entry_count": 15,
      "failed_game_aliases": [
        "GB2-02"
      ],
      "first_pass_admitted": 3,
      "first_pass_rate": 0.1875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:42Z",
      "losses": 95,
      "max_score": 66.83613208021495,
      "measured_at": null,
      "min_score": 15.80923544929635,
      "model_code_failure_rate": 0.0625,
      "model_code_failures": 1,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": 0.9375,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 604,
      "rating_method": "sequential_elo",
      "raw_elo": 1661.675761419006,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 12,
      "repaired_rate": 0.75,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 50.191363917808594,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 222,
      "wins": 368
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1455,
      "capped_draws": 69,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.712055,
        "mean": 0.49864374999999994,
        "median": 0.5065824999999999,
        "min": 0.29423750000000004,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 3.98915
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 48.01757453308687,
      "contributing_entrants": null,
      "distinct_rated_opponents": 625,
      "draws": 388,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:45Z",
      "losses": 261,
      "max_score": 99.15475058347155,
      "measured_at": null,
      "min_score": 14.741184593734893,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1277,
      "rating_method": "sequential_elo",
      "raw_elo": 1644.2583933799935,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 48.01757453308687,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 178,
      "wins": 628
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1561,
      "capped_draws": 232,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.58673,
        "mean": 0.442856875,
        "median": 0.4176025,
        "min": 0.34269,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 3.5428550000000003
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 45.05538252246229,
      "contributing_entrants": null,
      "distinct_rated_opponents": 652,
      "draws": 481,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:44:30Z",
      "losses": 268,
      "max_score": 81.04306968418157,
      "measured_at": null,
      "min_score": 20.085766120064036,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1359,
      "rating_method": "sequential_elo",
      "raw_elo": 1584.536956534744,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 45.05538252246229,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 199,
      "wins": 610
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1403,
      "capped_draws": 170,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.34174499999999997,
        "mean": 0.217940625,
        "median": 0.19638,
        "min": 0.14813,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 1.743525
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 44.07615889983954,
      "contributing_entrants": null,
      "distinct_rated_opponents": 657,
      "draws": 491,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 8,
      "first_pass_rate": 1.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:39:26Z",
      "losses": 322,
      "max_score": 91.0317105229074,
      "measured_at": null,
      "min_score": 11.892584996054628,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": 1.0,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1387,
      "rating_method": "sequential_elo",
      "raw_elo": 1606.1167124424544,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 44.07615889983954,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 8,
      "wins": 574
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1211,
      "capped_draws": 144,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 2,
        "max": 0.072978,
        "mean": 0.048872,
        "median": 0.046325000000000005,
        "min": 0.03731,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 0.390976
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 44.64232956611945,
      "contributing_entrants": null,
      "distinct_rated_opponents": 571,
      "draws": 291,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:00Z",
      "losses": 402,
      "max_score": 64.12287600481088,
      "measured_at": null,
      "min_score": 21.75260522744053,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": 0.75,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1209,
      "rating_method": "sequential_elo",
      "raw_elo": 1558.5916787756134,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 41.54436881586808,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 1,
      "wins": 516
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1203,
      "capped_draws": 25,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 18
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.226689264,
        "mean": 0.16037539425,
        "median": 0.140728509,
        "min": 0.10689655,
        "recorded_attempt_count": 18,
        "run_count": 8,
        "total": 1.283003154
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 43.5273049102168,
      "contributing_entrants": null,
      "distinct_rated_opponents": 554,
      "draws": 309,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-06"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T14:35:59Z",
      "losses": 361,
      "max_score": 59.41026994182745,
      "measured_at": null,
      "min_score": 17.581567025013598,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": 0.75,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1201,
      "rating_method": "sequential_elo",
      "raw_elo": 1596.5699667703077,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.75,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 40.50672145306643,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 531
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1453,
      "capped_draws": 146,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.046574,
        "mean": 0.036155625,
        "median": 0.035338,
        "min": 0.025803,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 9,
        "token_priced_attempt_count": 9,
        "total": 0.315048
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 40.92861802674063,
      "contributing_entrants": null,
      "distinct_rated_opponents": 645,
      "draws": 442,
      "entry_count": 8,
      "failed_game_aliases": [
        "GB2-06"
      ],
      "first_pass_admitted": 8,
      "first_pass_rate": 0.8888888888888888,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:14Z",
      "losses": 345,
      "max_score": 73.91132063132736,
      "measured_at": null,
      "min_score": 11.282582150463083,
      "model_code_failure_rate": 0.1111111111111111,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": 0.8888888888888888,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1273,
      "rating_method": "sequential_elo",
      "raw_elo": 1549.3497212912257,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 39.74101455866921,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 177,
      "wins": 486
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 965,
      "capped_draws": 128,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 17
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.2996388,
        "mean": 0.13471628689999998,
        "median": 0.1144847257,
        "min": 0.0276132876,
        "recorded_attempt_count": 17,
        "run_count": 8,
        "total": 1.0777302952
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 42.4596419290968,
      "contributing_entrants": null,
      "distinct_rated_opponents": 460,
      "draws": 247,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:39:33Z",
      "losses": 429,
      "max_score": 68.412551151382,
      "measured_at": null,
      "min_score": 13.36353157637584,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": 0.75,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 965,
      "rating_method": "sequential_elo",
      "raw_elo": 1523.4278723519426,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.75,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 39.51314909495272,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 492
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1329,
      "capped_draws": 188,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.309345,
        "mean": 0.239705,
        "median": 0.232815,
        "min": 0.175235,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 1.91764
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 38.86135552659715,
      "contributing_entrants": null,
      "distinct_rated_opponents": 629,
      "draws": 429,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 8,
      "first_pass_rate": 1.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:38Z",
      "losses": 399,
      "max_score": 61.54507370770128,
      "measured_at": null,
      "min_score": 14.385271315724854,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": 1.0,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1304,
      "rating_method": "sequential_elo",
      "raw_elo": 1527.9629146538246,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 38.86135552659715,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 16,
      "wins": 476
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1198,
      "capped_draws": 181,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 16
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.08827968,
        "mean": 0.05437612,
        "median": 0.0574344,
        "min": 0.00733824,
        "recorded_attempt_count": 16,
        "run_count": 8,
        "total": 0.43500896
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 41.3850467207755,
      "contributing_entrants": null,
      "distinct_rated_opponents": 552,
      "draws": 367,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-04"
      ],
      "first_pass_admitted": 1,
      "first_pass_rate": 0.125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T15:12:19Z",
      "losses": 378,
      "max_score": 78.90361697494383,
      "measured_at": null,
      "min_score": 12.24169582676053,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": 0.75,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1198,
      "rating_method": "sequential_elo",
      "raw_elo": 1531.8791047817433,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 38.51312557252109,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 453
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1226,
      "capped_draws": 147,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.21477375,
        "mean": 0.17756265624999998,
        "median": 0.17285124999999998,
        "min": 0.1418975,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 1.4205012499999998
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 39.65661326048428,
      "contributing_entrants": null,
      "distinct_rated_opponents": 585,
      "draws": 279,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-02"
      ],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:37Z",
      "losses": 465,
      "max_score": 68.30589565602956,
      "measured_at": null,
      "min_score": 15.994049148715748,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": 0.875,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1223,
      "rating_method": "sequential_elo",
      "raw_elo": 1504.5443847461988,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 38.35461566711227,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 479
    },
    {
      "admitted_plugins": 16,
      "attempted_matches": 2576,
      "capped_draws": 394,
      "codegen_cost_usd": {
        "admitted_cell_count": 16,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 46
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 8,
        "max": 0.16378006679999998,
        "mean": 0.12038942953333334,
        "median": 0.1261480173333333,
        "min": 0.05546312440000001,
        "recorded_attempt_count": 46,
        "run_count": 24,
        "total": 2.8893463088
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": 41.131076513149246,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1213,
      "draws": 723,
      "entry_count": 16,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06"
      ],
      "first_pass_admitted": 3,
      "first_pass_rate": 0.125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:57Z",
      "losses": 624,
      "max_score": 64.67108012218564,
      "measured_at": null,
      "min_score": 17.591072904112846,
      "model_code_failure_rate": 0.3333333333333333,
      "model_code_failures": 8,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": 0.6666666666666666,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 2342,
      "rating_method": "sequential_elo",
      "raw_elo": 1558.5901142740408,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 13,
      "repaired_rate": 0.5416666666666666,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 37.166123147911485,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 212,
      "wins": 995
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1510,
      "capped_draws": 158,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 11
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.24673750000000003,
        "mean": 0.16744,
        "median": 0.158095,
        "min": 0.108735,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 11
        },
        "run_count": 8,
        "token_priced_attempt_count": 11,
        "total": 1.33952
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 35.910461569583475,
      "contributing_entrants": null,
      "distinct_rated_opponents": 660,
      "draws": 471,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:31Z",
      "losses": 373,
      "max_score": 66.13437229960988,
      "measured_at": null,
      "min_score": 10.838789505864996,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1306,
      "rating_method": "sequential_elo",
      "raw_elo": 1506.2791701522935,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.910461569583475,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 195,
      "wins": 462
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1368,
      "capped_draws": 170,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.132112,
        "mean": 0.0780535,
        "median": 0.06797,
        "min": 0.05673,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 0.624428
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 36.75508899972182,
      "contributing_entrants": null,
      "distinct_rated_opponents": 637,
      "draws": 458,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:07Z",
      "losses": 419,
      "max_score": 66.50406286570629,
      "measured_at": null,
      "min_score": 10.164730563488815,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": 0.875,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1364,
      "rating_method": "sequential_elo",
      "raw_elo": 1509.2425222105776,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.54835364117075,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 487
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1378,
      "capped_draws": 179,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 15
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.121855,
        "mean": 0.09110296875,
        "median": 0.09280125,
        "min": 0.0343775,
        "recorded_attempt_count": 15,
        "run_count": 8,
        "total": 0.7288237499999999
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 36.66582431412571,
      "contributing_entrants": null,
      "distinct_rated_opponents": 642,
      "draws": 464,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 1,
      "first_pass_rate": 0.125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T14:28:19Z",
      "losses": 435,
      "max_score": 76.51799749682782,
      "measured_at": null,
      "min_score": 12.808440532491185,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": 0.875,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1377,
      "rating_method": "sequential_elo",
      "raw_elo": 1514.0501560831765,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.75,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.4620196749746,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 478
    },
    {
      "admitted_plugins": 25,
      "attempted_matches": 4507,
      "capped_draws": 360,
      "codegen_cost_usd": {
        "admitted_cell_count": 25,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 49
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 7,
        "max": 0.0362005625,
        "mean": 0.024352715625,
        "median": 0.025206018750000003,
        "min": 0.016582062499999998,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 49
        },
        "run_count": 32,
        "token_priced_attempt_count": 49,
        "total": 0.7792869
      },
      "codegen_evaluated_cells": 32,
      "conditional_success_score": 34.0794804876644,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1981,
      "draws": 1101,
      "entry_count": 25,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-08"
      ],
      "first_pass_admitted": 13,
      "first_pass_rate": 0.40625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:14Z",
      "losses": 1856,
      "max_score": 51.745308794671715,
      "measured_at": null,
      "min_score": 12.825845883037182,
      "model_code_failure_rate": 0.21875,
      "model_code_failures": 7,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": 0.78125,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 4097,
      "rating_method": "sequential_elo",
      "raw_elo": 1412.1309477956484,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 12,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 32.03984993317203,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 402,
      "wins": 1140
    },
    {
      "admitted_plugins": 23,
      "attempted_matches": 3833,
      "capped_draws": 420,
      "codegen_cost_usd": {
        "admitted_cell_count": 23,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 26
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.29749325000000004,
        "mean": 0.24155631250000004,
        "median": 0.24021375,
        "min": 0.17873725,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 26
        },
        "run_count": 24,
        "token_priced_attempt_count": 26,
        "total": 5.7973514999999995
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": 31.71867304290317,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1659,
      "draws": 1313,
      "entry_count": 23,
      "failed_game_aliases": [
        "GB2-06"
      ],
      "first_pass_admitted": 22,
      "first_pass_rate": 0.9166666666666666,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:12Z",
      "losses": 1050,
      "max_score": 41.874942556165,
      "measured_at": null,
      "min_score": 13.393525783812322,
      "model_code_failure_rate": 0.041666666666666664,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": 0.9583333333333334,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 3327,
      "rating_method": "sequential_elo",
      "raw_elo": 1476.9048931117868,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.041666666666666664,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 31.38297846515569,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 492,
      "wins": 964
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 770,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 4,
        "cell_count": 7,
        "cost_basis_counts": {
          "provider_reported": 15
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.0092087,
        "mean": 0.006571171428571429,
        "median": 0.00666192,
        "min": 0.00456754,
        "recorded_attempt_count": 15,
        "run_count": 7,
        "total": 0.0459982
      },
      "codegen_evaluated_cells": 7,
      "conditional_success_score": 34.3314236075538,
      "contributing_entrants": null,
      "distinct_rated_opponents": 362,
      "draws": 135,
      "entry_count": 4,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-08"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01"
      ],
      "infra_failures_excluded": 1,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:48Z",
      "losses": 379,
      "max_score": 58.56858588745748,
      "measured_at": null,
      "min_score": 19.078764110307976,
      "model_code_failure_rate": 0.42857142857142855,
      "model_code_failures": 3,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": 0.5714285714285714,
      "provider": "Google",
      "provider_key": "google",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 770,
      "rating_method": "sequential_elo",
      "raw_elo": 1391.9757267092311,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 4,
      "repaired_rate": 0.5714285714285714,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 29.8491728115773,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 256
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1389,
      "capped_draws": 171,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 17
        },
        "evidence_basis": "Provider-reported",
        "max": 0.07439860000000001,
        "mean": 0.060660574999999994,
        "median": 0.0565846,
        "min": 0.0485456,
        "recorded_attempt_count": 17,
        "run_count": 8,
        "total": 0.4852846
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 29.208909638100376,
      "contributing_entrants": null,
      "distinct_rated_opponents": 651,
      "draws": 485,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T14:16:21Z",
      "losses": 540,
      "max_score": 63.82915309593504,
      "measured_at": null,
      "min_score": 12.039456124515215,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": 1.0,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1377,
      "rating_method": "sequential_elo",
      "raw_elo": 1412.3182837229035,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 8,
      "repaired_rate": 1.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 29.208909638100376,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 3,
      "wins": 352
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 976,
      "capped_draws": 178,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 19
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 4,
        "max": 0.0445032,
        "mean": 0.03204184375,
        "median": 0.0312573,
        "min": 0.0200767,
        "recorded_attempt_count": 19,
        "run_count": 9,
        "total": 0.2854256
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 30.924385903086723,
      "contributing_entrants": null,
      "distinct_rated_opponents": 455,
      "draws": 358,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-04",
        "GB2-05"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:59:37Z",
      "losses": 384,
      "max_score": 62.91507461001133,
      "measured_at": null,
      "min_score": 8.615240570387764,
      "model_code_failure_rate": 0.4444444444444444,
      "model_code_failures": 4,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": 0.5555555555555556,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 976,
      "rating_method": "sequential_elo",
      "raw_elo": 1419.8732201685202,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.5555555555555556,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 26.69825998643895,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 234
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1006,
      "capped_draws": 133,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 15
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.0,
        "mean": 0.0,
        "median": 0.0,
        "min": 0.0,
        "recorded_attempt_count": 15,
        "run_count": 8,
        "total": 0.0
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 27.523696618442827,
      "contributing_entrants": null,
      "distinct_rated_opponents": 480,
      "draws": 263,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02"
      ],
      "first_pass_admitted": 3,
      "first_pass_rate": 0.375,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:02:27Z",
      "losses": 470,
      "max_score": 48.887746203916265,
      "measured_at": null,
      "min_score": 13.352273105809124,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": 0.75,
      "provider": "Other",
      "provider_key": "other",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 987,
      "rating_method": "sequential_elo",
      "raw_elo": 1380.91863340719,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 25.61368581357492,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 4,
      "wins": 254
    },
    {
      "admitted_plugins": 14,
      "attempted_matches": 2520,
      "capped_draws": 301,
      "codegen_cost_usd": {
        "admitted_cell_count": 14,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 23
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.004375912,
        "mean": 0.00311882929125,
        "median": 0.002947833025,
        "min": 0.001831717,
        "recorded_attempt_count": 23,
        "run_count": 16,
        "total": 0.04990126866
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": 22.27339080380156,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1070,
      "draws": 770,
      "entry_count": 14,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-04"
      ],
      "first_pass_admitted": 10,
      "first_pass_rate": 0.625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:57Z",
      "losses": 1102,
      "max_score": 39.83364255519466,
      "measured_at": null,
      "min_score": 7.826698848192246,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 2,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": 0.875,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 2192,
      "rating_method": "sequential_elo",
      "raw_elo": 1361.0509111428591,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 4,
      "repaired_rate": 0.25,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 21.542115517324167,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 321,
      "wins": 320
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1189,
      "capped_draws": 176,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 16
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.00444719,
        "mean": 0.0024656575,
        "median": 0.002155255,
        "min": 0.001130258,
        "recorded_attempt_count": 16,
        "run_count": 8,
        "total": 0.01972526
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 21.673949582168344,
      "contributing_entrants": null,
      "distinct_rated_opponents": 560,
      "draws": 496,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-07"
      ],
      "first_pass_admitted": 1,
      "first_pass_rate": 0.125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:06:15Z",
      "losses": 482,
      "max_score": 27.909510548557073,
      "measured_at": null,
      "min_score": 14.80141330228754,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": 0.875,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1173,
      "rating_method": "sequential_elo",
      "raw_elo": 1358.4089995121935,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 6,
      "repaired_rate": 0.75,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 20.962355023916732,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 11,
      "wins": 195
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1346,
      "capped_draws": 144,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 17
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.09705768,
        "mean": 0.0730637925,
        "median": 0.07051098,
        "min": 0.05225028,
        "recorded_attempt_count": 17,
        "run_count": 8,
        "total": 0.58451034
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 20.939560331755704,
      "contributing_entrants": null,
      "distinct_rated_opponents": 627,
      "draws": 426,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-13T11:16:26Z",
      "losses": 675,
      "max_score": 61.310936966259646,
      "measured_at": null,
      "min_score": 3.993926331153074,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": 0.875,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1346,
      "rating_method": "sequential_elo",
      "raw_elo": 1328.1272666706718,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 7,
      "repaired_rate": 0.875,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 20.25207708705361,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 245
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 673,
      "capped_draws": 6,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 17
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.13491350000000002,
        "mean": 0.10247140625,
        "median": 0.113468875,
        "min": 0.061553,
        "recorded_attempt_count": 17,
        "run_count": 8,
        "total": 0.81977125
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 22.564812642546553,
      "contributing_entrants": null,
      "distinct_rated_opponents": 338,
      "draws": 26,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-04"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T12:51:58Z",
      "losses": 488,
      "max_score": 36.75902784191273,
      "measured_at": null,
      "min_score": 2.4688491519502116,
      "model_code_failure_rate": 0.375,
      "model_code_failures": 3,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": 0.625,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 665,
      "rating_method": "sequential_elo",
      "raw_elo": 1267.7455324044884,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.625,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 20.063270856813254,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 4,
      "wins": 151
    },
    {
      "admitted_plugins": 15,
      "attempted_matches": 2937,
      "capped_draws": 237,
      "codegen_cost_usd": {
        "admitted_cell_count": 15,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 46
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 6,
        "max": 0.02904086,
        "mean": 0.020196066154166664,
        "median": 0.0210709803,
        "min": 0.0085050301,
        "recorded_attempt_count": 46,
        "run_count": 21,
        "total": 0.426437947
      },
      "codegen_evaluated_cells": 21,
      "conditional_success_score": 20.319806819135895,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1351,
      "draws": 1103,
      "entry_count": 15,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-06",
        "GB2-08"
      ],
      "first_pass_admitted": 1,
      "first_pass_rate": 0.047619047619047616,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-03",
        "GB2-06",
        "GB2-07"
      ],
      "infra_failures_excluded": 3,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:44:12Z",
      "losses": 1237,
      "max_score": 31.688441494119953,
      "measured_at": null,
      "min_score": 12.99127610042693,
      "model_code_failure_rate": 0.2857142857142857,
      "model_code_failures": 6,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": 0.7142857142857143,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 2913,
      "rating_method": "sequential_elo",
      "raw_elo": 1341.463192009627,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 14,
      "repaired_rate": 0.6666666666666666,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 18.680459977813825,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 24,
      "wins": 573
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1176,
      "capped_draws": 178,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 16
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.0149315,
        "mean": 0.00915140625,
        "median": 0.00943475,
        "min": 0.00402875,
        "recorded_attempt_count": 16,
        "run_count": 8,
        "total": 0.07321125
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 19.47051412152977,
      "contributing_entrants": null,
      "distinct_rated_opponents": 554,
      "draws": 300,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03"
      ],
      "first_pass_admitted": 2,
      "first_pass_rate": 0.25,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:21Z",
      "losses": 677,
      "max_score": 32.62660804381708,
      "measured_at": null,
      "min_score": 7.287827335921403,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": 0.75,
      "provider": "Google",
      "provider_key": "google",
      "rank": 30,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1175,
      "rating_method": "sequential_elo",
      "raw_elo": 1242.0219913838712,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 4,
      "repaired_rate": 0.5,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 18.11935505071165,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 198
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 765,
      "capped_draws": 0,
      "codegen_cost_usd": {
        "admitted_cell_count": 4,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 17
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 4,
        "max": 0.0711619,
        "mean": 0.0371769625,
        "median": 0.029592175,
        "min": 0.02387735,
        "recorded_attempt_count": 17,
        "run_count": 8,
        "total": 0.2974157
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 27.42111661651889,
      "contributing_entrants": null,
      "distinct_rated_opponents": 358,
      "draws": 307,
      "entry_count": 4,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-07",
        "GB2-08"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:33:38Z",
      "losses": 284,
      "max_score": 45.1976284044549,
      "measured_at": null,
      "min_score": 11.328963024338488,
      "model_code_failure_rate": 0.5,
      "model_code_failures": 4,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": 0.5,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": null,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "provisional",
      "rated_matches": 765,
      "rating_method": "sequential_elo",
      "raw_elo": 1423.2064040998605,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": 4,
      "repaired_rate": 0.5,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 23.0583186650848,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 174
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 2098.912558734133,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 90.50823856284876,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 191
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 24,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1968.039713684342,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.05797906734983,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 178
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 229,
      "capped_draws": 46,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 46,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:01Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1817.5279719116684,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 68.05592761865948,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 145
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 13,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 27,
      "max_score": null,
      "measured_at": "2026-07-19T22:54:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1816.9440144348723,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 67.83390511372896,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 136
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 4,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 38,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1769.7396233879429,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 64.06461183077373,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 134
    },
    {
      "admitted_plugins": 0,
      "attempted_matches": 0,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 0,
      "draws": 20,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 36,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 0,
      "rating_method": "sequential_elo",
      "raw_elo": 1766.7949819491082,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 63.99399190662281,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 147
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 104,
      "capped_draws": 6,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 6,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 19,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1657.929362886318,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 52.54906076651749,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 24,
      "wins": 55
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 49,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 49,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 64,
      "max_score": null,
      "measured_at": "2026-07-18T07:45:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 208,
      "rating_method": "sequential_elo",
      "raw_elo": 1577.0187486349464,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.86718004327302,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 95
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 233,
      "capped_draws": 142,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 142,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 209,
      "rating_method": "sequential_elo",
      "raw_elo": 1574.5951623324984,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.67892214642133,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 53
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 48,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 48,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 50,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1571.8654129749273,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.29093722140246,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 82
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 27,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 27,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 75,
      "max_score": null,
      "measured_at": "2026-07-18T07:26:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 1532.9014110845703,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 45.333741935710094,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 104
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 3,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 3,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 106,
      "max_score": null,
      "measured_at": "2026-07-18T07:26:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 1532.363238012958,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 45.29076855762162,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 97
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 34,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 34,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 96,
      "max_score": null,
      "measured_at": "2026-07-18T13:13:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 216,
      "rating_method": "sequential_elo",
      "raw_elo": 1492.7658388863824,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.18065252242629,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 86
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 40,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 40,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 98,
      "max_score": null,
      "measured_at": "2026-07-18T13:20:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 216,
      "rating_method": "sequential_elo",
      "raw_elo": 1481.1080097227068,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 41.24976926801048,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 78
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 512,
      "capped_draws": 48,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 225,
      "draws": 48,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 170,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 461,
      "rating_method": "sequential_elo",
      "raw_elo": 1472.9202679788602,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 39.63678818057228,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 11.445626981434346,
      "unilateral_faults": 51,
      "wins": 243
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 38,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 38,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 80,
      "max_score": null,
      "measured_at": "2026-07-18T07:22:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1443.1087668140876,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 38.158374372380486,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 87
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 38,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 38,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 97,
      "max_score": null,
      "measured_at": "2026-07-13T09:10:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 204,
      "rating_method": "sequential_elo",
      "raw_elo": 1418.2850887948575,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.17077061768282,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 69
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 106,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 106,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 68,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 204,
      "rating_method": "sequential_elo",
      "raw_elo": 1409.0038878685773,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 35.429662268792015,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 30
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 416,
      "capped_draws": 140,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 184,
      "draws": 140,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 156,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 388,
      "rating_method": "sequential_elo",
      "raw_elo": 1392.0961122348049,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.01864551691267,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 92
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 95,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 95,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 75,
      "max_score": null,
      "measured_at": "2026-07-13T10:33:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 204,
      "rating_method": "sequential_elo",
      "raw_elo": 1373.9001010765544,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.62660804381708,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 34
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 275,
      "capped_draws": 53,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 128,
      "draws": 53,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 151,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 246,
      "rating_method": "sequential_elo",
      "raw_elo": 1389.6805588572147,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.54055632230109,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 16.700837490522975,
      "unilateral_faults": 29,
      "wins": 42
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 408,
      "capped_draws": 44,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 180,
      "draws": 44,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 254,
      "max_score": null,
      "measured_at": "2026-07-13T11:05:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 408,
      "rating_method": "sequential_elo",
      "raw_elo": 1307.2991411331964,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.308489424231603,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 110
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 14,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 140,
      "max_score": null,
      "measured_at": "2026-07-13T10:03:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 204,
      "rating_method": "sequential_elo",
      "raw_elo": 1270.388251226152,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 24.36113714812227,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 50
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 43,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 43,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 119,
      "max_score": null,
      "measured_at": "2026-07-13T10:29:49Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 204,
      "rating_method": "sequential_elo",
      "raw_elo": 1244.695464286948,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.309555778625047,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 42
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 32,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 32,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 141,
      "max_score": null,
      "measured_at": "2026-07-13T10:43:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 204,
      "rating_method": "sequential_elo",
      "raw_elo": 1228.7411778838975,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.03559835136842,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 31
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 204,
      "capped_draws": 52,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 52,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 131,
      "max_score": null,
      "measured_at": "2026-07-13T10:33:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 204,
      "rating_method": "sequential_elo",
      "raw_elo": 1193.6662700102047,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.23485012188387,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 21
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 863,
      "capped_draws": 141,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 360,
      "draws": 141,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 571,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 791,
      "rating_method": "sequential_elo",
      "raw_elo": 1126.4084745532457,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.825845883037182,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 79
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 78,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 2,
      "max_score": null,
      "measured_at": "2026-07-13T08:41:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1734.0416529644124,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.803509565086735,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 108
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 212,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 126,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-19T22:54:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1628.0873090440657,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.58888031678649,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 25,
      "wins": 57
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 40,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 23,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 0,
      "max_score": null,
      "measured_at": "2026-07-18T13:32:58Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 40,
      "rating_method": "sequential_elo",
      "raw_elo": 1624.3920370977314,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.234252268553153,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 17
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 89,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1592.071710504841,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.03133313019584,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 67
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 158,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 7,
      "max_score": null,
      "measured_at": "2026-07-13T04:24:13Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1526.4347319128728,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.13978047432457,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 23
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 190,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 158,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 20,
      "max_score": null,
      "measured_at": "2026-07-13T03:10:13Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1511.4661034845751,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.581567025013598,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 11
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 197,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 174,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-18T13:06:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 197,
      "rating_method": "sequential_elo",
      "raw_elo": 1494.6898896066327,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.80141330228754,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 9
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 220,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 166,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 19,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 195,
      "rating_method": "sequential_elo",
      "raw_elo": 1494.4839505766154,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.741184593734893,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 25,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 194,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 85,
      "draws": 161,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 9,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 170,
      "rating_method": "sequential_elo",
      "raw_elo": 1495.1147491406418,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.502998442395736,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 0
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 197,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 178,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 9,
      "max_score": null,
      "measured_at": "2026-07-18T13:18:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 197,
      "rating_method": "sequential_elo",
      "raw_elo": 1492.2674296240918,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.385271315724854,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 10
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 650,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 252,
      "draws": 520,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 42,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 578,
      "rating_method": "sequential_elo",
      "raw_elo": 1486.8152333470407,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 13.393525783812322,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 16
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 571,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 248,
      "draws": 507,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-18T13:29:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 571,
      "rating_method": "sequential_elo",
      "raw_elo": 1484.647053718562,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.99127610042693,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 21
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 164,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-13T03:19:46Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1483.757284789138,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.808440532491185,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 11
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 619,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 248,
      "draws": 516,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 571,
      "rating_method": "sequential_elo",
      "raw_elo": 1480.6048343694315,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.29688395269576,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 179,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-13T03:20:50Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1480.3814624068068,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.24169582676053,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 2
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 170,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-13T02:44:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1479.204177726602,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.039456124515215,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 174,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-13T03:26:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1478.4258751778857,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.892584996054628,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 176,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-13T03:18:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1475.0682309617343,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.328963024338488,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 150,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1477.1096050255824,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.282582150463083,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 1
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 151,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1474.526184085617,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.838789505864996,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 172,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-13T03:17:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1468.3676297864029,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.164730563488815,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 173,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 15,
      "max_score": null,
      "measured_at": "2026-07-13T03:02:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1459.2710168841108,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 8.615240570387764,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 1
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 430,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 166,
      "draws": 331,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 44,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 381,
      "rating_method": "sequential_elo",
      "raw_elo": 1454.5728435719077,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.826698848192246,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 163,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 22,
      "max_score": null,
      "measured_at": "2026-07-13T03:49:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1452.4666259599655,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.433178600457437,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 55,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 3,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1614.322264306601,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 65.27953913368336,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 27,
      "wins": 13
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 39,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 21,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 0,
      "max_score": null,
      "measured_at": "2026-07-18T07:22:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 22,
      "rating_method": "sequential_elo",
      "raw_elo": 1592.817950871737,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.72016331680182,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 60.5072070285374,
      "unilateral_faults": 13,
      "wins": 8
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 142,
      "capped_draws": 14,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 53,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1543.8491195243553,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 53.46761933203254,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 54,
      "wins": 21
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 55,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 6,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 22,
      "rating_method": "sequential_elo",
      "raw_elo": 1560.1206345087253,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 52.26076337429936,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 60.5072070285374,
      "unilateral_faults": 27,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 38,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-15T06:47:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1535.4586566733617,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 44.876066552892205,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 9,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 54,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 2,
      "max_score": null,
      "measured_at": "2026-07-19T22:51:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 21,
      "rating_method": "sequential_elo",
      "raw_elo": 1524.8474314032603,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.649960732537416,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 62.850358177805205,
      "unilateral_faults": 30,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 51,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 33,
      "draws": 25,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-18T13:37:53Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 48,
      "rating_method": "sequential_elo",
      "raw_elo": 1487.3909516384024,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 40.238950200339005,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 27.67857142857143,
      "unilateral_faults": 0,
      "wins": 10
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 174,
      "capped_draws": 21,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 64,
      "draws": 40,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 70,
      "rating_method": "sequential_elo",
      "raw_elo": 1507.5695867541574,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 39.17797027985697,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 58.02799666097794,
      "unilateral_faults": 90,
      "wins": 16
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 61,
      "capped_draws": 6,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 16,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-19T22:51:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 25,
      "rating_method": "sequential_elo",
      "raw_elo": 1499.142262426009,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.76759726733822,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 54.30806756909202,
      "unilateral_faults": 33,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 31,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 9,
      "max_score": null,
      "measured_at": "2026-07-18T12:51:58Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 24,
      "rating_method": "sequential_elo",
      "raw_elo": 1496.7974806168354,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.75902784191273,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 56.25,
      "unilateral_faults": 3,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 51,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 24,
      "rating_method": "sequential_elo",
      "raw_elo": 1487.841275971651,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.44189220241967,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 56.25,
      "unilateral_faults": 27,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 39,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-18T15:14:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 24,
      "rating_method": "sequential_elo",
      "raw_elo": 1486.690833169341,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.14425139093331,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 56.25,
      "unilateral_faults": 7,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 40,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-18T14:16:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 28,
      "rating_method": "sequential_elo",
      "raw_elo": 1468.8713045965055,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.0143622071467,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 49.09766908852594,
      "unilateral_faults": 3,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 52,
      "capped_draws": 6,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 21,
      "draws": 15,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 22,
      "rating_method": "sequential_elo",
      "raw_elo": 1471.901085116667,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 29.436735776150815,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 60.5072070285374,
      "unilateral_faults": 29,
      "wins": 2
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 43,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-14T04:35:01Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 24,
      "rating_method": "sequential_elo",
      "raw_elo": 1458.2668032208471,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 26.790429608055035,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 56.25,
      "unilateral_faults": 4,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 61,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 17,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 25,
      "rating_method": "sequential_elo",
      "raw_elo": 1455.028170391493,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 26.35446649301918,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 54.30806756909202,
      "unilateral_faults": 27,
      "wins": 2
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 36,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-15T06:47:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1439.1506653901242,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.95940957471972,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 11,
      "wins": 1
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 56,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1383.947621293607,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 5.677361994071518,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 28,
      "wins": 1
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 95,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-13T07:20:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1612.691053164444,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 44.00470514710234,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 55
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 372,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 162,
      "draws": 201,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 46,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 324,
      "rating_method": "sequential_elo",
      "raw_elo": 1548.2603105218825,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.934062470661942,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 77
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 162,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 97,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 26,
      "max_score": null,
      "measured_at": "2026-07-13T05:55:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1538.5360765579255,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 30.114997119927704,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 39
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 571,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 250,
      "draws": 340,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 49,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 499,
      "rating_method": "sequential_elo",
      "raw_elo": 1526.3772493624135,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.91403849157441,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 110
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 130,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 15,
      "max_score": null,
      "measured_at": "2026-07-13T06:10:45Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1524.2028580866609,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.469379833230263,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 19
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 133,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 18,
      "max_score": null,
      "measured_at": "2026-07-13T05:29:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1518.4707885990597,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 26.39710935603434,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 13
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 112,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 18,
      "max_score": null,
      "measured_at": "2026-07-13T05:24:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1515.9469791946428,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.907259785377427,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 33
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 133,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 18,
      "max_score": null,
      "measured_at": "2026-07-13T05:22:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1514.954893807325,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.72167517689021,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 125,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-19T22:50:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1513.6327304771148,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.456448781739248,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 20
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 135,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 21,
      "max_score": null,
      "measured_at": "2026-07-13T05:34:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1512.569092518357,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.293107653611095,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 130,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1505.2245904025017,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 24.120137012879244,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 30
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 109,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-18T13:32:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1503.0333541118373,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.710233052691397,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 37
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 127,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 32,
      "max_score": null,
      "measured_at": "2026-07-18T13:30:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 175,
      "rating_method": "sequential_elo",
      "raw_elo": 1493.0370519838302,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.824317816622944,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 16
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 133,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 28,
      "max_score": null,
      "measured_at": "2026-07-18T13:30:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 175,
      "rating_method": "sequential_elo",
      "raw_elo": 1492.6536957756052,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.75260522744053,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 14
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 135,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-13T04:49:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1486.637632962378,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.44223510703427,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 107,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 37,
      "max_score": null,
      "measured_at": "2026-07-18T13:30:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1485.004398644425,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.337643672154105,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 32
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 198,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 87,
      "draws": 133,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 20,
      "max_score": null,
      "measured_at": "2026-07-19T22:39:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 174,
      "rating_method": "sequential_elo",
      "raw_elo": 1484.1160491905316,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.13941705311589,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 21
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 110,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 29,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1484.9225058954553,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.085766120064036,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 23
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 98,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 37,
      "max_score": null,
      "measured_at": "2026-07-13T04:47:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1481.5462303723896,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.489811083464367,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 29
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 348,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 162,
      "draws": 139,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 90,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:16Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 324,
      "rating_method": "sequential_elo",
      "raw_elo": 1478.8569618553822,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.9511141734381,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 95
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 65,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 42,
      "draws": 51,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-08T14:07:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 58,
      "rating_method": "sequential_elo",
      "raw_elo": 1488.0734362229514,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.7945013339251,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 21.344455490411935,
      "unilateral_faults": 7,
      "wins": 2
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 499,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 250,
      "draws": 353,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 106,
      "max_score": null,
      "measured_at": "2026-07-18T13:32:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 499,
      "rating_method": "sequential_elo",
      "raw_elo": 1464.180007111608,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.279702862860805,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 40
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 131,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 23,
      "max_score": null,
      "measured_at": "2026-07-19T23:11:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1463.894901142338,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.152233940860192,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 8
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 104,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 55,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 14,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1493.2004102426245,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.80923544929635,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 24,
      "wins": 11
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 128,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 25,
      "max_score": null,
      "measured_at": "2026-07-19T22:49:41Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1461.904495198417,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.779898340965923,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 9
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 118,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 31,
      "max_score": null,
      "measured_at": "2026-07-13T05:16:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1461.749475059943,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.768795178999987,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 14
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 119,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 29,
      "max_score": null,
      "measured_at": "2026-07-13T05:57:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1439.4513471639775,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.597592388925714,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 15
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 122,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 29,
      "max_score": null,
      "measured_at": "2026-07-13T05:58:01Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1432.7007067367829,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.334782779254471,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 247,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 24,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1944.2879504833952,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 81.04306968418157,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 199
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 28,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 1940.196351882442,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.56575826299661,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 164
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 41,
      "max_score": null,
      "measured_at": "2026-07-13T10:55:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1832.865700872228,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 72.40678999216624,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 182
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 64,
      "max_score": null,
      "measured_at": "2026-07-13T10:52:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1779.9574013908557,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 68.30589565602956,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 159
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1796.7384081258213,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 66.83613208021495,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 48,
      "wins": 72
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 74,
      "max_score": null,
      "measured_at": "2026-07-13T11:06:41Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1756.7107870985228,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 66.50406286570629,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 149
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 1754.0076749994753,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 66.13437229960988,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 144
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 76,
      "max_score": null,
      "measured_at": "2026-07-13T10:35:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1722.2000400310835,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 63.82915309593504,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 147
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 77,
      "max_score": null,
      "measured_at": "2026-07-13T09:23:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1677.9082554601025,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.39612027754948,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 146
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 73,
      "max_score": null,
      "measured_at": "2026-07-13T10:32:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1665.189160421109,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 59.41026994182745,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 150
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 72,
      "max_score": null,
      "measured_at": "2026-07-13T10:52:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1664.051984424949,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 59.322128031247175,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 151
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 83,
      "max_score": null,
      "measured_at": "2026-07-13T10:00:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1654.3300481477468,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.56858588745748,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 140
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 225,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 100,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 88,
      "max_score": null,
      "measured_at": "2026-07-18T07:30:17Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 225,
      "rating_method": "sequential_elo",
      "raw_elo": 1652.485643981534,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.43481351627299,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 137
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 228,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 102,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 99,
      "max_score": null,
      "measured_at": "2026-07-18T07:45:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 228,
      "rating_method": "sequential_elo",
      "raw_elo": 1619.4103608287369,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.88471911706098,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 129
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 274,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 143,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 80,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 241,
      "rating_method": "sequential_elo",
      "raw_elo": 1628.9910986264856,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.2402064105688,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 15.813918072220453,
      "unilateral_faults": 29,
      "wins": 161
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 198,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 79,
      "max_score": null,
      "measured_at": "2026-07-18T13:28:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1587.209662701708,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 53.23984773672601,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 119
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 233,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 104,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 108,
      "max_score": null,
      "measured_at": "2026-07-18T14:28:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 233,
      "rating_method": "sequential_elo",
      "raw_elo": 1556.9745692654642,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.067360858185296,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 125
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 255,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 103,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 109,
      "max_score": null,
      "measured_at": "2026-07-19T22:47:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 231,
      "rating_method": "sequential_elo",
      "raw_elo": 1553.0673307548009,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 50.755795487647504,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 122
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 233,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 104,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 92,
      "max_score": null,
      "measured_at": "2026-07-18T15:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 233,
      "rating_method": "sequential_elo",
      "raw_elo": 1551.8649996217848,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 50.67132081499249,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 141
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 229,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 102,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 111,
      "max_score": null,
      "measured_at": "2026-07-18T13:02:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 229,
      "rating_method": "sequential_elo",
      "raw_elo": 1529.080346867643,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.887746203916265,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 118
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 235,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 105,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 124,
      "max_score": null,
      "measured_at": "2026-07-18T13:33:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 235,
      "rating_method": "sequential_elo",
      "raw_elo": 1481.1343053572973,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 45.1976284044549,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 111
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 736,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 334,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 373,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 683,
      "rating_method": "sequential_elo",
      "raw_elo": 1434.9229171350576,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 40.87409094944158,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 7.9069590361102255,
      "unilateral_faults": 53,
      "wins": 310
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 543,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 238,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 295,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 490,
      "rating_method": "sequential_elo",
      "raw_elo": 1383.068198538091,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.67726968699771,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 10.542612048146967,
      "unilateral_faults": 53,
      "wins": 195
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 145,
      "max_score": null,
      "measured_at": "2026-07-13T08:46:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1313.814207923553,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.17538303972932,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 78
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 446,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 198,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 293,
      "max_score": null,
      "measured_at": "2026-07-13T09:47:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 446,
      "rating_method": "sequential_elo",
      "raw_elo": 1307.531858998345,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.688441494119953,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 153
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 139,
      "max_score": null,
      "measured_at": "2026-07-13T08:20:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1258.7774181863872,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.909510548557073,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 84
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 181,
      "max_score": null,
      "measured_at": "2026-07-13T10:29:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1168.0026818410374,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.873608699660835,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 42
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 496,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 198,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 350,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 448,
      "rating_method": "sequential_elo",
      "raw_elo": 1138.2211791864315,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.569853375981936,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 98
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 181,
      "max_score": null,
      "measured_at": "2026-07-13T10:25:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 1122.122660255199,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.317472469475717,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 42
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 210,
      "max_score": null,
      "measured_at": "2026-07-13T10:57:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 30,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 223,
      "rating_method": "sequential_elo",
      "raw_elo": 950.2269439894764,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 3.993926331153074,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 13
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 178,
      "rating_method": "sequential_elo",
      "raw_elo": 2104.6700311392274,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 99.15475058347155,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 164
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 2015.3034273186183,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 91.79025106882118,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 153
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 22,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:08Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 2006.0900644222068,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 91.0317105229074,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 153
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 2000.763555581476,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 90.58621353721448,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 151
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 20,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1943.2585453468528,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 85.85876169674393,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 149
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 386,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 182,
      "draws": 20,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 53,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 358,
      "rating_method": "sequential_elo",
      "raw_elo": 1882.2534328321856,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.84990486320305,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 285
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 7,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 34,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1858.78017083648,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 78.90361697494383,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 136
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 38,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1798.227461198194,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 73.91132063132736,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 130
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-13T11:05:20Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1731.3538953086795,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 68.412551151382,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 130
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 185,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 7,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 72,
      "max_score": null,
      "measured_at": "2026-07-18T13:59:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1663.928762676415,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.91507461001133,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 106
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 128,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 4,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 23,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1639.618492881366,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.185159863922095,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 48,
      "wins": 53
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 185,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 68,
      "max_score": null,
      "measured_at": "2026-07-18T12:51:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1603.7290733939203,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 57.958805581471466,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 103
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 259,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 136,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 97,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 230,
      "rating_method": "sequential_elo",
      "raw_elo": 1601.566834229231,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 56.460143022093256,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 16.700837490522975,
      "unilateral_faults": 29,
      "wins": 119
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 610,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 286,
      "draws": 16,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 234,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:24Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 557,
      "rating_method": "sequential_elo",
      "raw_elo": 1547.876116730321,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.745308794671715,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 22.47015523905305,
      "unilateral_faults": 53,
      "wins": 307
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 185,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 77,
      "max_score": null,
      "measured_at": "2026-07-18T13:33:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1524.5024971634189,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 51.43604390797115,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 94
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 7,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 93,
      "max_score": null,
      "measured_at": "2026-07-18T13:12:49Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1499.4169506198664,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 49.39629209571456,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 89
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 213,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 84,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1471.6275218884668,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 47.10837522132795,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 92
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 190,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 85,
      "max_score": null,
      "measured_at": "2026-07-18T13:32:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 190,
      "rating_method": "sequential_elo",
      "raw_elo": 1468.277296193658,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 46.83881208054259,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 96
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 181,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 4,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 97,
      "max_score": null,
      "measured_at": "2026-07-18T13:32:41Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 181,
      "rating_method": "sequential_elo",
      "raw_elo": 1417.0942451711392,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.56668046962967,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 80
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 201,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 115,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1264.153921743513,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 29.947755290899035,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 61
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 114,
      "max_score": null,
      "measured_at": "2026-07-13T10:53:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1235.9723508656575,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.627553167229117,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 63
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 135,
      "max_score": null,
      "measured_at": "2026-07-13T10:57:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Step 3.7 Flash",
      "model_slug": "stepfun-step-3-7-flash",
      "plugin_success_rate": null,
      "provider": "StepFun",
      "provider_key": "stepfun",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1212.42022199375,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "stepfun",
        "provider_model": "stepfun/step-3.7-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.68849520405192,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 41
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 130,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1161.8598348623989,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.525834502695627,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 47
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 138,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1160.555970204923,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.418486705783742,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 37
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 136,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:25Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1142.6674567245134,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.94571688942271,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 40
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 138,
      "max_score": null,
      "measured_at": "2026-07-13T11:13:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1132.1373003458395,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.078764110307976,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 37
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 6,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 131,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1121.5181057532589,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.204480772726864,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 40
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 132,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1121.2223397200323,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.180130214648468,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 33
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 405,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 181,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 293,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 357,
      "rating_method": "sequential_elo",
      "raw_elo": 1115.1927214246716,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.693979671992224,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 62
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 177,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 142,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 30,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 177,
      "rating_method": "sequential_elo",
      "raw_elo": 1062.5823635244633,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 13.352273105809124,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 34
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 207,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 26,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 207,
      "rating_method": "sequential_elo",
      "raw_elo": 1980.0726989542147,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 83.59400879681142,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 181
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 33,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1879.34321498302,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 76.51799749682782,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 172
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-20T08:39:26Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 1822.3096281620233,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 72.5215843938676,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 161
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 207,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 51,
      "max_score": null,
      "measured_at": "2026-07-20T08:39:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 183,
      "rating_method": "sequential_elo",
      "raw_elo": 1744.4392326074358,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 66.94034211955521,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 132
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 59,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 208,
      "rating_method": "sequential_elo",
      "raw_elo": 1724.8419077718663,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 65.69317630293021,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 149
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 108,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 42,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 84,
      "rating_method": "sequential_elo",
      "raw_elo": 1748.3221682231233,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 65.0021168384253,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 32.49928516630233,
      "unilateral_faults": 24,
      "wins": 68
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 476,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 226,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 122,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 441,
      "rating_method": "sequential_elo",
      "raw_elo": 1722.9560227490508,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 64.67108012218564,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 12.453570188460423,
      "unilateral_faults": 29,
      "wins": 319
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 210,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-20T08:44:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1708.9329446505674,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 64.46609074562843,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 138
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 207,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 62,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:00Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 207,
      "rating_method": "sequential_elo",
      "raw_elo": 1702.524761583259,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 64.12287600481088,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 145
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 209,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 55,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:45Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1683.9317117604437,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.70664219367879,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 130
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 75,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 1665.8468310037947,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 61.54507370770128,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 131
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 90,
      "max_score": null,
      "measured_at": "2026-07-13T10:30:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1662.57676750296,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 61.310936966259646,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 115
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 215,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 100,
      "max_score": null,
      "measured_at": "2026-07-18T14:35:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 215,
      "rating_method": "sequential_elo",
      "raw_elo": 1618.640015606257,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.27438692501393,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 115
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 92,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 216,
      "rating_method": "sequential_elo",
      "raw_elo": 1580.0717347851164,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.57306742942482,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 124
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 100,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 216,
      "rating_method": "sequential_elo",
      "raw_elo": 1579.284163468929,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.5178160525783,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 116
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 217,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 119,
      "max_score": null,
      "measured_at": "2026-07-18T13:39:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 217,
      "rating_method": "sequential_elo",
      "raw_elo": 1453.4926523128145,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 46.69739678054314,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 98
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 207,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 123,
      "max_score": null,
      "measured_at": "2026-07-18T07:33:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 207,
      "rating_method": "sequential_elo",
      "raw_elo": 1398.1757341379953,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.77153617992668,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 84
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 210,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 110,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1398.9471021180111,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.719305088473725,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 76
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 475,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 204,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 237,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 422,
      "rating_method": "sequential_elo",
      "raw_elo": 1407.821516156022,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 41.874942556165,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 24.700376732585976,
      "unilateral_faults": 53,
      "wins": 185
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 210,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 110,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1369.1231225975473,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 40.627029890432105,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 76
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 135,
      "max_score": null,
      "measured_at": "2026-07-13T10:26:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1357.5577991174084,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 39.91259802066606,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 70
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 250,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 107,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 125,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 201,
      "rating_method": "sequential_elo",
      "raw_elo": 1394.789298445333,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 39.83364255519466,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 42.67472243679027,
      "unilateral_faults": 49,
      "wins": 76
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 425,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 192,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 266,
      "max_score": null,
      "measured_at": "2026-07-20T08:44:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 425,
      "rating_method": "sequential_elo",
      "raw_elo": 1325.1285742366904,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.67131258724129,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 159
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 136,
      "max_score": null,
      "measured_at": "2026-07-13T10:26:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1281.0099027752608,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.54244730456053,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 69
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 140,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 1262.964769744279,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 33.2812350764154,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 66
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 642,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 290,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 417,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 590,
      "rating_method": "sequential_elo",
      "raw_elo": 1248.2717844825058,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 30.87267632350711,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 21.337361218395134,
      "unilateral_faults": 52,
      "wins": 173
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 165,
      "max_score": null,
      "measured_at": "2026-07-13T11:02:13Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 1109.86793342262,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.540860428737904,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 41
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 182,
      "max_score": null,
      "measured_at": "2026-07-13T11:13:17Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1019.198467476494,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.175295051675466,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 23
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 206,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 190,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 206,
      "rating_method": "sequential_elo",
      "raw_elo": 892.4461779045539,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.287827335921403,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 16
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-13T11:04:20Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Fable 5",
      "model_slug": "claude-fable-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1963.0657653404032,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-fable-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.70760851946686,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 178
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 212,
      "capped_draws": 14,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 20,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:26Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1939.7834663198096,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 78.19758392078708,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 154
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 193,
      "capped_draws": 21,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 21,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-18T13:59:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Kimi K2.7 Code",
      "model_slug": "moonshotai-kimi-k2-7-code",
      "plugin_success_rate": null,
      "provider": "Moonshot",
      "provider_key": "moonshotai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 193,
      "rating_method": "sequential_elo",
      "raw_elo": 1758.7242311131176,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "moonshotai",
        "provider_model": "moonshotai/kimi-k2.7-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.78251919520124,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 142
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 80,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 40,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 9,
      "max_score": null,
      "measured_at": "2026-07-18T14:36:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok 4.5",
      "model_slug": "x-ai-grok-4-5",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 80,
      "rating_method": "sequential_elo",
      "raw_elo": 1789.3560923707078,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-4.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.44733474350692,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 34.33688094430304,
      "unilateral_faults": 0,
      "wins": 71
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 218,
      "capped_draws": 89,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 89,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 32,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 194,
      "rating_method": "sequential_elo",
      "raw_elo": 1546.4684267251985,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 35.982615143180006,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 73
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 609,
      "capped_draws": 219,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 281,
      "draws": 219,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 161,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 561,
      "rating_method": "sequential_elo",
      "raw_elo": 1526.7555232791372,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 33.8075748539808,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 181
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 78,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 102,
      "draws": 78,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-18T13:18:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1514.225764389215,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.579257820966184,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 79
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 104,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 102,
      "draws": 104,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 49,
      "max_score": null,
      "measured_at": "2026-07-18T14:36:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "O3",
      "model_slug": "o3",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 201,
      "rating_method": "sequential_elo",
      "raw_elo": 1498.275464427157,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "o3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 30.85786947221065,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 1,
      "wins": 48
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 89,
      "capped_draws": 23,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 53,
      "draws": 23,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 38,
      "max_score": null,
      "measured_at": "2026-07-12T07:18:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 83,
      "rating_method": "sequential_elo",
      "raw_elo": 1507.320284523039,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 30.164266572203793,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 10.804472558998102,
      "unilateral_faults": 6,
      "wins": 22
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 222,
      "capped_draws": 109,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 109,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 47,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1469.93922509029,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.79030487189539,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 42
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 83,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 102,
      "draws": 83,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 72,
      "max_score": null,
      "measured_at": "2026-07-18T13:36:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1469.102467467738,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.73055959801432,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 47
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 649,
      "capped_draws": 351,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 290,
      "draws": 351,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 146,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 577,
      "rating_method": "sequential_elo",
      "raw_elo": 1429.1002585284762,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.356176321343483,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 80
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 154,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 154,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 34,
      "max_score": null,
      "measured_at": "2026-07-18T07:43:08Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1409.7887856234097,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.247306751579977,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 129,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 129,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-18T07:38:53Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-OSS 120B",
      "model_slug": "openai-gpt-oss-120b",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1398.424496391982,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "openai/gpt-oss-120b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.026163511594184,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 14
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 132,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 132,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-13T10:37:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1398.105571299315,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.000131987943323,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 9
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 127,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 127,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-13T11:06:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1395.8762814960626,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.73567193586623,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 14
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 132,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 132,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:41Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1383.3975126241307,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.411447512045243,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 11
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 411,
      "capped_draws": 193,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 194,
      "draws": 193,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 144,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 387,
      "rating_method": "sequential_elo",
      "raw_elo": 1379.937603115728,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.08326171470216,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 50
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 129,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 129,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 51,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1379.9491051706354,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.049139240086024,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 9
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 129,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 129,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:26Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1379.1134741276585,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.93443399794906,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 9
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 430,
      "capped_draws": 248,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 210,
      "draws": 248,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 109,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 398,
      "rating_method": "sequential_elo",
      "raw_elo": 1396.5514516371084,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.591072904112846,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 23.65262231176029,
      "unilateral_faults": 29,
      "wins": 41
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 132,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 132,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 55,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:46Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1375.6570776322492,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.57139943394984,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 423,
      "capped_draws": 247,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 189,
      "draws": 247,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 110,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 375,
      "rating_method": "sequential_elo",
      "raw_elo": 1371.9351520062514,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.175614049147185,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 18
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 115,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 115,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 66,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 5",
      "model_slug": "gpt-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1361.0557279723228,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.994049148715748,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 126,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 126,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 60,
      "max_score": null,
      "measured_at": "2026-07-13T11:13:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.5 Flash",
      "model_slug": "google-gemini-3-5-flash",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1357.7971814937282,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.5-flash",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.64390388899189,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 212,
      "capped_draws": 127,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 127,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 51,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 26,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1355.313167227012,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.39366014033035,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 132,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 132,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 47,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Grok Build 0.1",
      "model_slug": "x-ai-grok-build-0-1",
      "plugin_success_rate": null,
      "provider": "xAI",
      "provider_key": "x-ai",
      "rank": 27,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1349.190480538579,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "x-ai",
        "provider_model": "x-ai/grok-build-0.1",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.72744628666598,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 128,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 128,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 58,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 28,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1336.5754200595743,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 13.36353157637584,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 37,
      "capped_draws": 5,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 36,
      "draws": 5,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-08T14:06:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 29,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 36,
      "rating_method": "sequential_elo",
      "raw_elo": 1271.6952570692756,
      "reasoning_level": "Medium",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "medium",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/medium"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "medium",
        "reasoning_bucket_label": "Medium reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "medium"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 2.4688491519502116,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/medium",
      "track_name": "Medium variants",
      "uncertainty": 38.449493652678655,
      "unilateral_faults": 1,
      "wins": 1
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1599,
      "capped_draws": 127,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 10
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.1598575,
        "mean": 0.093565,
        "median": 0.07461,
        "min": 0.06826499999999999,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 10
        },
        "run_count": 8,
        "token_priced_attempt_count": 10,
        "total": 0.7485199999999999
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 48.68896152233137,
      "contributing_entrants": null,
      "distinct_rated_opponents": 657,
      "draws": 446,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 6,
      "first_pass_rate": 0.75,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:24Z",
      "losses": 293,
      "max_score": 91.35944335010576,
      "measured_at": null,
      "min_score": 11.529384168381274,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1385,
      "rating_method": "sequential_elo",
      "raw_elo": 1642.005780104476,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.25,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 48.68896152233137,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 204,
      "wins": 646
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1407,
      "capped_draws": 130,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.144235,
        "mean": 0.12174125,
        "median": 0.121475,
        "min": 0.09594,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 0.97393
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 45.67057735948555,
      "contributing_entrants": null,
      "distinct_rated_opponents": 660,
      "draws": 413,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 8,
      "first_pass_rate": 1.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:59Z",
      "losses": 388,
      "max_score": 78.90896540288884,
      "measured_at": null,
      "min_score": 11.482980464184529,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": 1.0,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1391,
      "rating_method": "sequential_elo",
      "raw_elo": 1573.4311153668432,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 45.67057735948555,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 7,
      "wins": 590
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1366,
      "capped_draws": 41,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.190125,
        "mean": 0.15785312499999998,
        "median": 0.15388000000000002,
        "min": 0.131495,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 1.262825
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 44.86404683281706,
      "contributing_entrants": null,
      "distinct_rated_opponents": 638,
      "draws": 335,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:52:02Z",
      "losses": 435,
      "max_score": 80.76792349183744,
      "measured_at": null,
      "min_score": 13.717353446889966,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": 0.875,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1366,
      "rating_method": "sequential_elo",
      "raw_elo": 1583.4197099575754,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 43.391079874656214,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 596
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1492,
      "capped_draws": 79,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.147405,
        "mean": 0.121464375,
        "median": 0.124865,
        "min": 0.09045,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 9,
        "token_priced_attempt_count": 9,
        "total": 1.0694850000000002
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 40.136337833904406,
      "contributing_entrants": null,
      "distinct_rated_opponents": 653,
      "draws": 348,
      "entry_count": 8,
      "failed_game_aliases": [
        "GB2-04"
      ],
      "first_pass_admitted": 8,
      "first_pass_rate": 0.8888888888888888,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:29Z",
      "losses": 405,
      "max_score": 61.398138957666426,
      "measured_at": null,
      "min_score": 14.58766401287794,
      "model_code_failure_rate": 0.1111111111111111,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": 0.8888888888888888,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1288,
      "rating_method": "sequential_elo",
      "raw_elo": 1528.1411119357276,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 38.971723529651854,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 196,
      "wins": 535
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1054,
      "capped_draws": 61,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 13
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.09803328,
        "mean": 0.0253809975,
        "median": 0.00943521,
        "min": 0.00639384,
        "recorded_attempt_count": 13,
        "run_count": 8,
        "total": 0.20304798
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 39.60055944956829,
      "contributing_entrants": null,
      "distinct_rated_opponents": 486,
      "draws": 238,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-04",
        "GB2-06"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.5,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:59:28Z",
      "losses": 412,
      "max_score": 54.04543152948865,
      "measured_at": null,
      "min_score": 6.05124929753971,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": 0.75,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1041,
      "rating_method": "sequential_elo",
      "raw_elo": 1517.3290306459585,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.25,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 36.852473046929816,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 4,
      "wins": 391
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1129,
      "capped_draws": 179,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 9
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.03912625,
        "mean": 0.022799843750000003,
        "median": 0.0216425,
        "min": 0.015285,
        "recorded_attempt_count": 9,
        "run_count": 8,
        "total": 0.18239875
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 37.35756527528258,
      "contributing_entrants": null,
      "distinct_rated_opponents": 538,
      "draws": 275,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-04"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T12:57:38Z",
      "losses": 379,
      "max_score": 70.14174273877752,
      "measured_at": null,
      "min_score": 11.919704676069871,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": 0.875,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1109,
      "rating_method": "sequential_elo",
      "raw_elo": 1500.1511607756418,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 36.13104954225295,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 10,
      "wins": 455
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1205,
      "capped_draws": 218,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 12
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.029078938,
        "mean": 0.015562441375,
        "median": 0.0131167185,
        "min": 0.005612805,
        "recorded_attempt_count": 12,
        "run_count": 8,
        "total": 0.124499531
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 36.77193379881386,
      "contributing_entrants": null,
      "distinct_rated_opponents": 562,
      "draws": 404,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-04"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.5,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:22:40Z",
      "losses": 396,
      "max_score": 75.05378217838995,
      "measured_at": null,
      "min_score": 10.465061755242804,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": 0.875,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1189,
      "rating_method": "sequential_elo",
      "raw_elo": 1491.2315536146675,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.56464539535866,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 7,
      "wins": 389
    },
    {
      "admitted_plugins": 28,
      "attempted_matches": 4908,
      "capped_draws": 240,
      "codegen_cost_usd": {
        "admitted_cell_count": 28,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 1,
          "provider_reported": 47
        },
        "evidence_basis": "Mixed: Provider-reported + Standard list-price estimate",
        "failed_cell_count": 4,
        "max": 0.0370864994,
        "mean": 0.02414278511875,
        "median": 0.021930847225,
        "min": 0.0140238704,
        "recorded_attempt_count": 47,
        "run_count": 32,
        "token_priced_attempt_count": 1,
        "total": 0.7725691238000001
      },
      "codegen_evaluated_cells": 32,
      "conditional_success_score": 36.77177511539484,
      "contributing_entrants": null,
      "distinct_rated_opponents": 2263,
      "draws": 1243,
      "entry_count": 28,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-07",
        "GB2-08"
      ],
      "first_pass_admitted": 16,
      "first_pass_rate": 0.5,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:44:44Z",
      "losses": 1634,
      "max_score": 47.55645299565777,
      "measured_at": null,
      "min_score": 23.89158662732705,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 4,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": 0.875,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 4661,
      "rating_method": "sequential_elo",
      "raw_elo": 1519.54710336448,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 12,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.56449192180031,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 218,
      "wins": 1784
    },
    {
      "admitted_plugins": 17,
      "attempted_matches": 3169,
      "capped_draws": 359,
      "codegen_cost_usd": {
        "admitted_cell_count": 17,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 35
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 7,
        "max": 0.004834030033333333,
        "mean": 0.0027724950333333333,
        "median": 0.0022439992166666664,
        "min": 0.00169122,
        "recorded_attempt_count": 35,
        "run_count": 24,
        "total": 0.06653988079999999
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": 38.57796942583274,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1471,
      "draws": 782,
      "entry_count": 17,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-07",
        "GB2-08"
      ],
      "first_pass_admitted": 12,
      "first_pass_rate": 0.5,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:38:40Z",
      "losses": 1468,
      "max_score": 39.07689302053589,
      "measured_at": null,
      "min_score": 37.704641705021,
      "model_code_failure_rate": 0.2916666666666667,
      "model_code_failures": 7,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": 0.7083333333333334,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 3131,
      "rating_method": "sequential_elo",
      "raw_elo": 1397.376338033076,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.20833333333333337,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 35.39148490151478,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 35,
      "wins": 881
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1412,
      "capped_draws": 167,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_list_price_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.07806,
        "mean": 0.06121625,
        "median": 0.061152,
        "min": 0.049788,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 0.48973
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 34.30827338646576,
      "contributing_entrants": null,
      "distinct_rated_opponents": 659,
      "draws": 487,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 8,
      "first_pass_rate": 1.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01"
      ],
      "infra_failures_excluded": 1,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:44:37Z",
      "losses": 404,
      "max_score": 60.31532930836284,
      "measured_at": null,
      "min_score": 11.593620745823015,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": 1.0,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1392,
      "rating_method": "sequential_elo",
      "raw_elo": 1520.2481939112397,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 34.30827338646576,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 5,
      "wins": 501
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1407,
      "capped_draws": 134,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 11
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.094594,
        "mean": 0.0571955,
        "median": 0.043012,
        "min": 0.032534,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 11
        },
        "run_count": 8,
        "token_priced_attempt_count": 11,
        "total": 0.457564
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 34.16293015239342,
      "contributing_entrants": null,
      "distinct_rated_opponents": 664,
      "draws": 445,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.625,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:41Z",
      "losses": 511,
      "max_score": 57.80317048612072,
      "measured_at": null,
      "min_score": 8.667504440488187,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1396,
      "rating_method": "sequential_elo",
      "raw_elo": 1474.9462497663255,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 3,
      "repaired_rate": 0.375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 34.16293015239342,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 3,
      "wins": 440
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1489,
      "capped_draws": 188,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 10
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.1192425,
        "mean": 0.0683559375,
        "median": 0.06152125,
        "min": 0.0542775,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 10
        },
        "run_count": 9,
        "token_priced_attempt_count": 10,
        "total": 0.6063375
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 34.051070978069944,
      "contributing_entrants": null,
      "distinct_rated_opponents": 651,
      "draws": 499,
      "entry_count": 8,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.7777777777777778,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:09Z",
      "losses": 434,
      "max_score": 54.32908419338588,
      "measured_at": null,
      "min_score": 14.61450799184122,
      "model_code_failure_rate": 0.1111111111111111,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": 0.8888888888888888,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1284,
      "rating_method": "sequential_elo",
      "raw_elo": 1463.699438330303,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.1111111111111111,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 33.06302955535,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 198,
      "wins": 351
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1043,
      "capped_draws": 52,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 7,
        "cost_basis_counts": {
          "provider_reported": 10
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.1311545862,
        "mean": 0.02955606902857143,
        "median": 0.0063034062,
        "min": 0.0057139812,
        "recorded_attempt_count": 10,
        "run_count": 7,
        "total": 0.2068924832
      },
      "codegen_evaluated_cells": 7,
      "conditional_success_score": 34.32361051437528,
      "contributing_entrants": null,
      "distinct_rated_opponents": 478,
      "draws": 215,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-04"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.5714285714285714,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T14:36:19Z",
      "losses": 469,
      "max_score": 55.17575767805614,
      "measured_at": null,
      "min_score": 11.938705607217443,
      "model_code_failure_rate": 0.14285714285714285,
      "model_code_failures": 1,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": 0.8571428571428571,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1030,
      "rating_method": "sequential_elo",
      "raw_elo": 1437.4398464646738,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.2857142857142857,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 33.02602214581218,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 3,
      "wins": 346
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1228,
      "capped_draws": 162,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 13
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.12956125,
        "mean": 0.031130218749999997,
        "median": 0.0158481,
        "min": 0.0111232,
        "recorded_attempt_count": 13,
        "run_count": 9,
        "total": 0.378603
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 34.8799485598801,
      "contributing_entrants": null,
      "distinct_rated_opponents": 576,
      "draws": 309,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-07"
      ],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.5555555555555556,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:43:56Z",
      "losses": 521,
      "max_score": 63.965994099933845,
      "measured_at": null,
      "min_score": 11.126777189882008,
      "model_code_failure_rate": 0.2222222222222222,
      "model_code_failures": 2,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": 0.7777777777777778,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1188,
      "rating_method": "sequential_elo",
      "raw_elo": 1389.945061414994,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.2222222222222222,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 32.75591371384714,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 36,
      "wins": 358
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1378,
      "capped_draws": 282,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 8
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 1,
        "max": 0.156555,
        "mean": 0.1422525,
        "median": 0.14850249999999998,
        "min": 0.12078,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 8
        },
        "run_count": 8,
        "token_priced_attempt_count": 8,
        "total": 1.13802
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 31.54347880987911,
      "contributing_entrants": null,
      "distinct_rated_opponents": 571,
      "draws": 590,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-06"
      ],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:41Z",
      "losses": 311,
      "max_score": 54.800844026442405,
      "measured_at": null,
      "min_score": 9.392722470573146,
      "model_code_failure_rate": 0.125,
      "model_code_failures": 1,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": 0.875,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1200,
      "rating_method": "sequential_elo",
      "raw_elo": 1482.789410802258,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 0,
      "repaired_rate": 0.0,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 30.507849941945324,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 177,
      "wins": 299
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1069,
      "capped_draws": 148,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 11
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 2,
        "max": 0.001632522,
        "mean": 0.00126739675,
        "median": 0.001232095,
        "min": 0.00090937,
        "recorded_attempt_count": 11,
        "run_count": 8,
        "total": 0.010139174
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 30.453962994232025,
      "contributing_entrants": null,
      "distinct_rated_opponents": 465,
      "draws": 418,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-06",
        "GB2-07"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.5,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-18T13:05:44Z",
      "losses": 286,
      "max_score": 77.08002467089872,
      "measured_at": null,
      "min_score": 0.0,
      "model_code_failure_rate": 0.25,
      "model_code_failures": 2,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": 0.75,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 984,
      "rating_method": "sequential_elo",
      "raw_elo": 1487.271601952845,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.25,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 28.34060594134785,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 3,
      "wins": 280
    },
    {
      "admitted_plugins": 24,
      "attempted_matches": 4122,
      "capped_draws": 465,
      "codegen_cost_usd": {
        "admitted_cell_count": 24,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 29
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.021615250000000003,
        "mean": 0.017288937500000004,
        "median": 0.01830825,
        "min": 0.01259225,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 29
        },
        "run_count": 24,
        "token_priced_attempt_count": 29,
        "total": 0.4149345
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": 25.12416565392437,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1766,
      "draws": 1370,
      "entry_count": 24,
      "failed_game_aliases": [],
      "first_pass_admitted": 19,
      "first_pass_rate": 0.7916666666666666,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:19Z",
      "losses": 1651,
      "max_score": 42.444850252935154,
      "measured_at": null,
      "min_score": 9.614328033579907,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 3610,
      "rating_method": "sequential_elo",
      "raw_elo": 1357.113877781704,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.20833333333333337,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 25.12416565392437,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 512,
      "wins": 589
    },
    {
      "admitted_plugins": 25,
      "attempted_matches": 4375,
      "capped_draws": 530,
      "codegen_cost_usd": {
        "admitted_cell_count": 25,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 56
        },
        "evidence_basis": "Standard list-price estimate",
        "failed_cell_count": 7,
        "max": 0.013223525,
        "mean": 0.0089983359375,
        "median": 0.00845865,
        "min": 0.0058062875,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 56
        },
        "run_count": 32,
        "token_priced_attempt_count": 56,
        "total": 0.28794675
      },
      "codegen_evaluated_cells": 32,
      "conditional_success_score": 24.57030339770677,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1929,
      "draws": 1424,
      "entry_count": 25,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07"
      ],
      "first_pass_admitted": 10,
      "first_pass_rate": 0.3125,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:50:48Z",
      "losses": 1571,
      "max_score": 36.40186495059713,
      "measured_at": null,
      "min_score": 7.616710049308469,
      "model_code_failure_rate": 0.21875,
      "model_code_failures": 7,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": 0.78125,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 3900,
      "rating_method": "sequential_elo",
      "raw_elo": 1387.1360974343463,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 15,
      "repaired_rate": 0.46875,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 23.0997897388718,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 449,
      "wins": 905
    },
    {
      "admitted_plugins": 19,
      "attempted_matches": 3412,
      "capped_draws": 288,
      "codegen_cost_usd": {
        "admitted_cell_count": 19,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 34
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 5,
        "max": 0.0027823742,
        "mean": 0.0022230800541666665,
        "median": 0.002112839,
        "min": 0.001825724,
        "recorded_attempt_count": 34,
        "run_count": 24,
        "total": 0.0533539213
      },
      "codegen_evaluated_cells": 24,
      "conditional_success_score": 23.474400757062273,
      "contributing_entrants": null,
      "distinct_rated_opponents": 1509,
      "draws": 1015,
      "entry_count": 19,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-06"
      ],
      "first_pass_admitted": 14,
      "first_pass_rate": 0.5833333333333334,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:10Z",
      "losses": 1507,
      "max_score": 33.96211011336977,
      "measured_at": null,
      "min_score": 11.9686486964846,
      "model_code_failure_rate": 0.20833333333333337,
      "model_code_failures": 5,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": 0.7916666666666666,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 3124,
      "rating_method": "sequential_elo",
      "raw_elo": 1363.3018849655905,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 5,
      "repaired_rate": 0.20833333333333337,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 22.142675954211537,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 279,
      "wins": 602
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1167,
      "capped_draws": 162,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 7,
        "cost_basis_counts": {
          "provider_reported": 8
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 1,
        "max": 0.00391754,
        "mean": 0.0023617857142857143,
        "median": 0.00186618,
        "min": 0.00144036,
        "recorded_attempt_count": 8,
        "run_count": 7,
        "total": 0.0165325
      },
      "codegen_evaluated_cells": 7,
      "conditional_success_score": 21.852753172641428,
      "contributing_entrants": null,
      "distinct_rated_opponents": 548,
      "draws": 288,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-03"
      ],
      "first_pass_admitted": 5,
      "first_pass_rate": 0.7142857142857143,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-02"
      ],
      "infra_failures_excluded": 1,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:44:35Z",
      "losses": 663,
      "max_score": 38.43379421672025,
      "measured_at": null,
      "min_score": 14.0470230789226,
      "model_code_failure_rate": 0.14285714285714285,
      "model_code_failures": 1,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": 0.8571428571428571,
      "provider": "Google",
      "provider_key": "google",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1167,
      "rating_method": "sequential_elo",
      "raw_elo": 1273.4990604635973,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.14285714285714285,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 21.0266198517886,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 216
    },
    {
      "admitted_plugins": 6,
      "attempted_matches": 1190,
      "capped_draws": 37,
      "codegen_cost_usd": {
        "admitted_cell_count": 6,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 13
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.0154064,
        "mean": 0.00874386,
        "median": 0.0064215999999999995,
        "min": 0.00524064,
        "recorded_attempt_count": 13,
        "run_count": 9,
        "total": 0.08145696000000001
      },
      "codegen_evaluated_cells": 9,
      "conditional_success_score": 18.367680108906587,
      "contributing_entrants": null,
      "distinct_rated_opponents": 537,
      "draws": 345,
      "entry_count": 6,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-03",
        "GB2-08"
      ],
      "first_pass_admitted": 4,
      "first_pass_rate": 0.4444444444444444,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-19T23:12:19Z",
      "losses": 636,
      "max_score": 29.091016958055572,
      "measured_at": null,
      "min_score": 7.661044648020888,
      "model_code_failure_rate": 0.3333333333333333,
      "model_code_failures": 3,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": 0.6666666666666666,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1165,
      "rating_method": "sequential_elo",
      "raw_elo": 1256.2197373900306,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.2222222222222222,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 16.597072548072685,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 24,
      "wins": 184
    },
    {
      "admitted_plugins": 8,
      "attempted_matches": 1368,
      "capped_draws": 164,
      "codegen_cost_usd": {
        "admitted_cell_count": 8,
        "cell_count": 8,
        "cost_basis_counts": {
          "local_standard_estimate": 9
        },
        "evidence_basis": "Standard list-price estimate",
        "max": 0.038815,
        "mean": 0.02196775,
        "median": 0.0196975,
        "min": 0.016152,
        "pricing_table_versions": {
          "boardgame-list-prices-2026-07-12": 9
        },
        "run_count": 8,
        "token_priced_attempt_count": 9,
        "total": 0.175742
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 12.591972433354108,
      "contributing_entrants": null,
      "distinct_rated_opponents": 639,
      "draws": 441,
      "entry_count": 8,
      "failed_game_aliases": [],
      "first_pass_admitted": 7,
      "first_pass_rate": 0.875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "infra_failures_excluded": 8,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:32Z",
      "losses": 738,
      "max_score": 29.22432622420127,
      "measured_at": null,
      "min_score": 0.0,
      "model_code_failure_rate": 0.0,
      "model_code_failures": 0,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": 1.0,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1259,
      "rating_method": "sequential_elo",
      "raw_elo": 1215.3309629673008,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 1,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 12.591972433354108,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 101,
      "wins": 80
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 1000,
      "capped_draws": 48,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 13
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 3,
        "max": 0.00870475,
        "mean": 0.005883593750000001,
        "median": 0.006313125,
        "min": 0.003432,
        "recorded_attempt_count": 13,
        "run_count": 8,
        "total": 0.04706875
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 12.522919483535407,
      "contributing_entrants": null,
      "distinct_rated_opponents": 453,
      "draws": 227,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-03",
        "GB2-04",
        "GB2-08"
      ],
      "first_pass_admitted": 3,
      "first_pass_rate": 0.375,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:43:51Z",
      "losses": 680,
      "max_score": 27.657743183844648,
      "measured_at": null,
      "min_score": 0.0,
      "model_code_failure_rate": 0.375,
      "model_code_failures": 3,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": 0.625,
      "provider": "Google",
      "provider_key": "google",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "ranked",
      "rated_matches": 1000,
      "rating_method": "sequential_elo",
      "raw_elo": 1128.5651786492597,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.25,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 11.13462493557314,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 0,
      "wins": 93
    },
    {
      "admitted_plugins": 7,
      "attempted_matches": 1287,
      "capped_draws": 150,
      "codegen_cost_usd": {
        "admitted_cell_count": 7,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 33
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 9,
        "max": 0.2156775,
        "mean": 0.157461515625,
        "median": 0.15705206249999998,
        "min": 0.098579625,
        "recorded_attempt_count": 33,
        "run_count": 16,
        "total": 2.51938425
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": 36.39955656915741,
      "contributing_entrants": null,
      "distinct_rated_opponents": 572,
      "draws": 433,
      "entry_count": 7,
      "failed_game_aliases": [
        "GB2-01",
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-07",
        "GB2-08"
      ],
      "first_pass_admitted": 0,
      "first_pass_rate": 0.0,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:39:29Z",
      "losses": 408,
      "max_score": 74.21628479825006,
      "measured_at": null,
      "min_score": 10.671164514073164,
      "model_code_failure_rate": 0.5625,
      "model_code_failures": 9,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": 0.4375,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": null,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "provisional",
      "rated_matches": 1111,
      "rating_method": "sequential_elo",
      "raw_elo": 1451.7221761896913,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 7,
      "repaired_rate": 0.4375,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 29.603332785792052,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 148,
      "wins": 270
    },
    {
      "admitted_plugins": 5,
      "attempted_matches": 1041,
      "capped_draws": 244,
      "codegen_cost_usd": {
        "admitted_cell_count": 5,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 29
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 11,
        "max": 0.0,
        "mean": 0.0,
        "median": 0.0,
        "min": 0.0,
        "recorded_attempt_count": 29,
        "run_count": 16,
        "total": 0.0
      },
      "codegen_evaluated_cells": 16,
      "conditional_success_score": 23.45410855826144,
      "contributing_entrants": null,
      "distinct_rated_opponents": 463,
      "draws": 244,
      "entry_count": 5,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-05",
        "GB2-06",
        "GB2-08"
      ],
      "first_pass_admitted": 3,
      "first_pass_rate": 0.1875,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-20T08:51:52Z",
      "losses": 568,
      "max_score": 25.92335810041868,
      "measured_at": null,
      "min_score": 20.9848590161042,
      "model_code_failure_rate": 0.6875,
      "model_code_failures": 11,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": 0.3125,
      "provider": "Other",
      "provider_key": "other",
      "rank": null,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "provisional",
      "rated_matches": 957,
      "rating_method": "sequential_elo",
      "raw_elo": 1258.2951271972113,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.125,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 17.53603632361322,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 72,
      "wins": 145
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 898,
      "capped_draws": 128,
      "codegen_cost_usd": {
        "admitted_cell_count": 4,
        "cell_count": 8,
        "cost_basis_counts": {
          "provider_reported": 14
        },
        "evidence_basis": "Provider-reported",
        "failed_cell_count": 4,
        "max": 0.06200699999999999,
        "mean": 0.0429946875,
        "median": 0.0423105,
        "min": 0.0223935,
        "recorded_attempt_count": 14,
        "run_count": 8,
        "total": 0.34395749999999997
      },
      "codegen_evaluated_cells": 8,
      "conditional_success_score": 18.10868187686674,
      "contributing_entrants": null,
      "distinct_rated_opponents": 364,
      "draws": 142,
      "entry_count": 4,
      "failed_game_aliases": [
        "GB2-02",
        "GB2-03",
        "GB2-04",
        "GB2-07"
      ],
      "first_pass_admitted": 2,
      "first_pass_rate": 0.25,
      "first_seen_at": null,
      "game_alias": null,
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": 0,
      "is_recent_addition": false,
      "last_measured_at": "2026-07-13T11:16:04Z",
      "losses": 477,
      "max_score": 23.811718728840177,
      "measured_at": null,
      "min_score": 11.269722122355963,
      "model_code_failure_rate": 0.5,
      "model_code_failures": 4,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": 0.5,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": null,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "provisional",
      "rated_matches": 772,
      "rating_method": "sequential_elo",
      "raw_elo": 1213.039686533759,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": 2,
      "repaired_rate": 0.25,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "track",
      "score": 15.227525675227149,
      "score_basis": "successful-game mean adjusted by the quarter power of plugin success rate",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": null,
      "unilateral_faults": 126,
      "wins": 153
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 234,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 203,
      "rating_method": "sequential_elo",
      "raw_elo": 2109.5036475112493,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 91.35944335010576,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 197
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 271,
      "capped_draws": 22,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 22,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1930.7455809922349,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 77.08002467089872,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 164
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 209,
      "capped_draws": 47,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 47,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 27,
      "max_score": null,
      "measured_at": "2026-07-13T11:01:25Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1746.3205149839891,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 62.35359462708452,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 128
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 31,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 31,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 40,
      "max_score": null,
      "measured_at": "2026-07-13T11:01:25Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1722.617410716957,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.47727375069463,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 134
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 41,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 41,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 42,
      "max_score": null,
      "measured_at": "2026-07-13T11:02:25Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1713.1783327478593,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 59.723558844154695,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 122
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 9,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 58,
      "max_score": null,
      "measured_at": "2026-07-13T11:04:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1689.1285291879904,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 57.80317048612072,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 138
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 209,
      "capped_draws": 36,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 36,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 62,
      "max_score": null,
      "measured_at": "2026-07-18T07:36:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 209,
      "rating_method": "sequential_elo",
      "raw_elo": 1628.8354772367124,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 53.01003732658696,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 111
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 21,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 21,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 60,
      "max_score": null,
      "measured_at": "2026-07-19T22:39:05Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 178,
      "rating_method": "sequential_elo",
      "raw_elo": 1630.8461985271176,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 52.98735499556829,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 97
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 209,
      "capped_draws": 29,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 29,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 62,
      "max_score": null,
      "measured_at": "2026-07-18T07:40:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 209,
      "rating_method": "sequential_elo",
      "raw_elo": 1616.6633385119806,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 52.03808620685494,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 118
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 57,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 57,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 44,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:46Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 178,
      "rating_method": "sequential_elo",
      "raw_elo": 1567.8460010354645,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 47.95675898609834,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 77
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 242,
      "capped_draws": 47,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 47,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 95,
      "max_score": null,
      "measured_at": "2026-07-19T23:11:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 218,
      "rating_method": "sequential_elo",
      "raw_elo": 1503.843655249275,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 43.07514483050649,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 76
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 241,
      "capped_draws": 141,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 141,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 35,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:45Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 217,
      "rating_method": "sequential_elo",
      "raw_elo": 1480.6476034854643,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 41.21798435610656,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 41
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 217,
      "capped_draws": 106,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 106,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 60,
      "max_score": null,
      "measured_at": "2026-07-18T13:26:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 217,
      "rating_method": "sequential_elo",
      "raw_elo": 1480.5938935798754,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 41.21369559436482,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 51
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 834,
      "capped_draws": 93,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 368,
      "draws": 93,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 384,
      "max_score": null,
      "measured_at": "2026-07-19T23:11:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 807,
      "rating_method": "sequential_elo",
      "raw_elo": 1465.608239185155,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 39.93290747538863,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 330
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 623,
      "capped_draws": 119,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 274,
      "draws": 119,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 255,
      "max_score": null,
      "measured_at": "2026-07-18T13:01:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 623,
      "rating_method": "sequential_elo",
      "raw_elo": 1452.8751682521522,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 38.95237355194132,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 249
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 217,
      "capped_draws": 33,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 33,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 102,
      "max_score": null,
      "measured_at": "2026-07-18T13:37:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 217,
      "rating_method": "sequential_elo",
      "raw_elo": 1445.7800593007478,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 38.43379421672025,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 82
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 701,
      "capped_draws": 127,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 316,
      "draws": 127,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 292,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 641,
      "rating_method": "sequential_elo",
      "raw_elo": 1429.5623071962452,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.40186495059713,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 7.9069590361102255,
      "unilateral_faults": 53,
      "wins": 222
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 509,
      "capped_draws": 105,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 222,
      "draws": 105,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 204,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 457,
      "rating_method": "sequential_elo",
      "raw_elo": 1429.831042124774,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.20768076040539,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 11.13389166034865,
      "unilateral_faults": 52,
      "wins": 148
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 66,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 66,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 84,
      "max_score": null,
      "measured_at": "2026-07-13T10:25:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1408.3063538759975,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 35.37938499429994,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 55
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 48,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 48,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 92,
      "max_score": null,
      "measured_at": "2026-07-13T10:42:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1395.7203702322447,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.37438815914073,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 65
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 308,
      "capped_draws": 10,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 124,
      "max_score": null,
      "measured_at": "2026-07-13T10:36:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 182,
      "rating_method": "sequential_elo",
      "raw_elo": 1247.9988882816424,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.44304861230847,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 126,
      "wins": 48
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 413,
      "capped_draws": 112,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 182,
      "draws": 112,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 219,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 385,
      "rating_method": "sequential_elo",
      "raw_elo": 1228.9690799556683,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.9848590161042,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 54
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 229,
      "capped_draws": 37,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 37,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 142,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1199.2151775483187,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.683354201876575,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 26
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 205,
      "capped_draws": 48,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 48,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 145,
      "max_score": null,
      "measured_at": "2026-07-13T10:55:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 205,
      "rating_method": "sequential_elo",
      "raw_elo": 1158.8381713068147,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.459230872743149,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 39,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 39,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-01",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 118,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:46Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 175,
      "rating_method": "sequential_elo",
      "raw_elo": 1156.9132723278346,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.123322831265655,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 18
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 195,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 170,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-18T13:32:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 195,
      "rating_method": "sequential_elo",
      "raw_elo": 1501.1789623177865,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.891286352323776,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 9
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 194,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 85,
      "draws": 153,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 170,
      "rating_method": "sequential_elo",
      "raw_elo": 1495.7638723484058,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.61450799184122,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 194,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 85,
      "draws": 144,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 12,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:39Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 170,
      "rating_method": "sequential_elo",
      "raw_elo": 1495.6076072588198,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.58766401287794,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 14
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 195,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 177,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-18T13:19:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 195,
      "rating_method": "sequential_elo",
      "raw_elo": 1491.0439817266465,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.150249973036471,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 193,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 86,
      "draws": 162,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 22,
      "max_score": null,
      "measured_at": "2026-07-18T13:27:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 193,
      "rating_method": "sequential_elo",
      "raw_elo": 1488.670887417819,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 13.717353446889966,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 9
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 153,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 24,
      "max_score": null,
      "measured_at": "2026-07-13T05:22:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1478.771633660387,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.938705607217443,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 97,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 54,
      "draws": 83,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 3,
      "max_score": null,
      "measured_at": "2026-07-08T14:06:26Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 92,
      "rating_method": "sequential_elo",
      "raw_elo": 1490.966094278961,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.919704676069871,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 8.09758473652126,
      "unilateral_faults": 5,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 211,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 171,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1476.3888776838344,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.529384168381274,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 158,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-13T03:35:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1476.1966495174229,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.482980464184529,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 15
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 124,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 31,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1473.5503997432495,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.671164514073164,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 7
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 168,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 15,
      "max_score": null,
      "measured_at": "2026-07-13T03:15:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1470.2711062713329,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.465061755242804,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 172,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 13,
      "max_score": null,
      "measured_at": "2026-07-13T02:42:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1468.472620449504,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 10.156109092208288,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 1
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 629,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 246,
      "draws": 512,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 41,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:39Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 557,
      "rating_method": "sequential_elo",
      "raw_elo": 1465.345170469226,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 9.614328033579907,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 211,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 174,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 12,
      "max_score": null,
      "measured_at": "2026-07-19T22:49:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 1463.9508694407311,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 9.392722470573146,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 1
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 162,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 150,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-13T03:30:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1462.0284427050033,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 8.691866558464978,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 1
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 395,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 164,
      "draws": 333,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 36,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 371,
      "rating_method": "sequential_elo",
      "raw_elo": 1459.5172262340852,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 8.610963018710688,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 2
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 185,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 169,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 16,
      "max_score": null,
      "measured_at": "2026-07-13T10:34:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1458.9800765394484,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 8.511944032346298,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 565,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 249,
      "draws": 465,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 61,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 541,
      "rating_method": "sequential_elo",
      "raw_elo": 1455.7451314903628,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.881138090855433,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 15
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 187,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 166,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 18,
      "max_score": null,
      "measured_at": "2026-07-13T03:32:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1453.9482658660377,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.661044648020888,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 2
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 791,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 327,
      "draws": 653,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 58,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 719,
      "rating_method": "sequential_elo",
      "raw_elo": 1454.2298523936438,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 7.616710049308469,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 168,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-13T03:20:30Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1444.5772700460002,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 6.05124929753971,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 1
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 160,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-02",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 26,
      "max_score": null,
      "measured_at": "2026-07-13T04:41:25Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1409.3515304162397,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 0.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 0
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 61,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 2,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 22,
      "rating_method": "sequential_elo",
      "raw_elo": 1544.1967765612917,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 48.14096707330006,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 60.5072070285374,
      "unilateral_faults": 36,
      "wins": 7
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 31,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 21,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-15T09:06:00Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 21,
      "rating_method": "sequential_elo",
      "raw_elo": 1535.1598787863509,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 45.31798142974239,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 62.850358177805205,
      "unilateral_faults": 3,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 38,
      "capped_draws": 5,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 21,
      "draws": 15,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 3,
      "max_score": null,
      "measured_at": "2026-07-18T15:14:36Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 23,
      "rating_method": "sequential_elo",
      "raw_elo": 1529.1953708321137,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 44.71416367512211,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 58.31207261596577,
      "unilateral_faults": 7,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 35,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 11,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-15T09:38:53Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 21,
      "rating_method": "sequential_elo",
      "raw_elo": 1527.6395418798847,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 43.37233130702697,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 62.850358177805205,
      "unilateral_faults": 11,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 33,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-15T09:38:53Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1527.705981575098,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.87030624668301,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 4,
      "wins": 6
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 159,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 65,
      "draws": 46,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 69,
      "rating_method": "sequential_elo",
      "raw_elo": 1520.5344462896417,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.444850252935154,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 58.45011939260174,
      "unilateral_faults": 90,
      "wins": 13
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 42,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-18T13:05:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 26,
      "rating_method": "sequential_elo",
      "raw_elo": 1506.35763614118,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 40.01373725333524,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 52.475044864937615,
      "unilateral_faults": 3,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 42,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 17,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-18T16:05:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 26,
      "rating_method": "sequential_elo",
      "raw_elo": 1501.979403510879,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 38.88100756843271,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 52.475044864937615,
      "unilateral_faults": 12,
      "wins": 4
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 85,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 41,
      "draws": 20,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 43,
      "rating_method": "sequential_elo",
      "raw_elo": 1502.6034117012573,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.13750839828127,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 61.6787826031713,
      "unilateral_faults": 33,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 30,
      "capped_draws": 5,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 21,
      "draws": 17,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 2,
      "max_score": null,
      "measured_at": "2026-07-18T12:57:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 22,
      "rating_method": "sequential_elo",
      "raw_elo": 1500.245554128479,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.76997367453777,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 60.5072070285374,
      "unilateral_faults": 5,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 44,
      "capped_draws": 6,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 7,
      "max_score": null,
      "measured_at": "2026-07-18T12:57:26Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 25,
      "rating_method": "sequential_elo",
      "raw_elo": 1493.9724275857818,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.43006546305724,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 54.30806756909202,
      "unilateral_faults": 5,
      "wins": 5
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 122,
      "capped_draws": 7,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 63,
      "draws": 32,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-19T22:54:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 66,
      "rating_method": "sequential_elo",
      "raw_elo": 1491.965718110616,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 34.59822313047601,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 60.65023878520347,
      "unilateral_faults": 45,
      "wins": 17
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 39,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 22,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 7,
      "max_score": null,
      "measured_at": "2026-07-15T09:56:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 26,
      "rating_method": "sequential_elo",
      "raw_elo": 1477.2621371920486,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.48619403221036,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 52.475044864937615,
      "unilateral_faults": 3,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 50,
      "capped_draws": 6,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 12,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1479.925857534619,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 30.50870524942116,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 28,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 57,
      "capped_draws": 3,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 6,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1476.7083357203906,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 29.67627290090365,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 30,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 54,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 7,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 8,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1474.9614690906308,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 29.22432622420127,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 29,
      "wins": 5
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 38,
      "capped_draws": 4,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 18,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 4,
      "max_score": null,
      "measured_at": "2026-07-14T04:26:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 23,
      "rating_method": "sequential_elo",
      "raw_elo": 1464.562698958926,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.99249714268673,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 58.31207261596577,
      "unilateral_faults": 4,
      "wins": 1
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 33,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 5,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-15T09:56:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 21,
      "rating_method": "sequential_elo",
      "raw_elo": 1464.8131949351737,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.11799441187583,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 62.850358177805205,
      "unilateral_faults": 3,
      "wins": 6
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 54,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 20,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 5,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 20,
      "rating_method": "sequential_elo",
      "raw_elo": 1460.6690821362936,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.52662161635416,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 65.3589451179962,
      "unilateral_faults": 33,
      "wins": 2
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 174,
      "capped_draws": 1,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 62,
      "draws": 27,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-03",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 35,
      "max_score": null,
      "measured_at": "2026-07-19T23:11:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 76,
      "rating_method": "sequential_elo",
      "raw_elo": 1424.8120002619244,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.6634480878402,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 53.69706000104055,
      "unilateral_faults": 79,
      "wins": 14
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 107,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 9,
      "max_score": null,
      "measured_at": "2026-07-13T06:51:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1569.6617737816275,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 35.95542625454721,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 47
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 115,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 17,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1541.5731286289524,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 30.68312374165008,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 30
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 100,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 20,
      "max_score": null,
      "measured_at": "2026-07-13T05:07:18Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1516.882072790436,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 26.082183215449735,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 43
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 198,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 87,
      "draws": 149,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 10,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 174,
      "rating_method": "sequential_elo",
      "raw_elo": 1510.0020478112715,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 24.981785442716276,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 15
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 176,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 133,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 25,
      "max_score": null,
      "measured_at": "2026-07-18T13:30:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1506.5058299871805,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 24.35981229662057,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 18
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 673,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 325,
      "draws": 488,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 79,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 649,
      "rating_method": "sequential_elo",
      "raw_elo": 1505.2434820044568,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.89158662732705,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 82
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 175,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 142,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 20,
      "max_score": null,
      "measured_at": "2026-07-18T13:22:34Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 175,
      "rating_method": "sequential_elo",
      "raw_elo": 1499.3418928953065,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.003733870007537,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 13
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 560,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 244,
      "draws": 338,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 87,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:39Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 488,
      "rating_method": "sequential_elo",
      "raw_elo": 1491.625368506487,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 21.35151432284935,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 63
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 138,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-13T04:51:47Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1486.6977379183295,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.45347865004077,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 15
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 199,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 126,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 175,
      "rating_method": "sequential_elo",
      "raw_elo": 1485.1438466765717,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.347774128566368,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 19
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 163,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 122,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 27,
      "max_score": null,
      "measured_at": "2026-07-13T04:50:17Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1483.2039916831088,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.782187694092556,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 14
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 122,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 23,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1482.305408162208,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.631826795717306,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 19
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 143,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 11,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1482.1741973198564,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.571653345574234,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 125,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 28,
      "max_score": null,
      "measured_at": "2026-07-13T04:45:16Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 164,
      "rating_method": "sequential_elo",
      "raw_elo": 1481.253860215171,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.435118814175468,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 11
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 324,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 162,
      "draws": 241,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-13T09:50:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 324,
      "rating_method": "sequential_elo",
      "raw_elo": 1476.5573768511383,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.520941944160917,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 35
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 162,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 81,
      "draws": 120,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 30,
      "max_score": null,
      "measured_at": "2026-07-13T04:47:20Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 162,
      "rating_method": "sequential_elo",
      "raw_elo": 1475.4230209775408,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.30874348485692,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 12
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 535,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 244,
      "draws": 374,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 89,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 487,
      "rating_method": "sequential_elo",
      "raw_elo": 1441.4986914364524,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.9686486964846,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 24
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 372,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 162,
      "draws": 213,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 82,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 324,
      "rating_method": "sequential_elo",
      "raw_elo": 1437.7913644144344,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.269171751144304,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 29
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 164,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 82,
      "draws": 121,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-04",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 32,
      "max_score": null,
      "measured_at": "2026-07-13T06:02:29Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 163,
      "rating_method": "sequential_elo",
      "raw_elo": 1423.787885302052,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 8.667504440488187,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 10
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 245,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 22,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1977.03183596768,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 83.57172066219728,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 199
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 59,
      "max_score": null,
      "measured_at": "2026-07-13T10:54:01Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1724.08563265649,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 63.965994099933845,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 162
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 54,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 1692.9024546427677,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 61.398138957666426,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 138
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 79,
      "max_score": null,
      "measured_at": "2026-07-13T10:41:24Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1659.2010072298715,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 58.93682106805347,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 142
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 249,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 83,
      "max_score": null,
      "measured_at": "2026-07-19T22:54:43Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 225,
      "rating_method": "sequential_elo",
      "raw_elo": 1605.601444995989,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 54.800844026442405,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 142
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 218,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 82,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 194,
      "rating_method": "sequential_elo",
      "raw_elo": 1601.5519559500542,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 54.32908419338588,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 112
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 229,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 101,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 103,
      "max_score": null,
      "measured_at": "2026-07-18T12:57:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 229,
      "rating_method": "sequential_elo",
      "raw_elo": 1530.6832284130016,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 49.01198470553688,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 126
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 233,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 103,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 130,
      "max_score": null,
      "measured_at": "2026-07-18T13:29:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 233,
      "rating_method": "sequential_elo",
      "raw_elo": 1493.3178966423554,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 46.133365799256566,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 103
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 236,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 106,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 124,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 208,
      "rating_method": "sequential_elo",
      "raw_elo": 1445.5223847352904,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 42.31096509029217,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 84
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 225,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 99,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 116,
      "max_score": null,
      "measured_at": "2026-07-18T13:05:46Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 225,
      "rating_method": "sequential_elo",
      "raw_elo": 1432.358546313694,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 41.37287849086039,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 109
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 664,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 291,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 395,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:27Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 664,
      "rating_method": "sequential_elo",
      "raw_elo": 1402.9550985696671,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 39.07689302053589,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 269
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 131,
      "max_score": null,
      "measured_at": "2026-07-13T09:11:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1375.0255732882074,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 36.91053334109965,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 90
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 902,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 400,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 513,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:09Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 874,
      "rating_method": "sequential_elo",
      "raw_elo": 1359.6807791452063,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 35.70615336320589,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 361
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 147,
      "max_score": null,
      "measured_at": "2026-07-13T09:09:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1327.4481272619323,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 33.22283063945276,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 74
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 724,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 298,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 440,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 676,
      "rating_method": "sequential_elo",
      "raw_elo": 1309.8287024578108,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.87673425923766,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 236
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 130,
      "max_score": null,
      "measured_at": "2026-07-13T08:06:26Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1274.1409172394135,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 29.091016958055572,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 91
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 152,
      "max_score": null,
      "measured_at": "2026-07-13T08:20:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1263.632103253458,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 28.276484338845368,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 69
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 222,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 153,
      "max_score": null,
      "measured_at": "2026-07-13T10:08:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 222,
      "rating_method": "sequential_elo",
      "raw_elo": 1205.969003299063,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.811718728840177,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 69
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 169,
      "max_score": null,
      "measured_at": "2026-07-13T10:39:33Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1190.26978308322,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.590209666987988,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 52
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 739,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 294,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 521,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 667,
      "rating_method": "sequential_elo",
      "raw_elo": 1109.4597446616813,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.332831857238713,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 146
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 483,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 232,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 392,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 454,
      "rating_method": "sequential_elo",
      "raw_elo": 1097.923684929318,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.481174430529984,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 11.13389166034865,
      "unilateral_faults": 29,
      "wins": 62
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 186,
      "max_score": null,
      "measured_at": "2026-07-13T11:01:16Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 1080.0485880274048,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.0470230789226,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 35
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 222,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 183,
      "max_score": null,
      "measured_at": "2026-07-13T10:40:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 222,
      "rating_method": "sequential_elo",
      "raw_elo": 1066.9209979297616,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 13.034181305418958,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 39
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 221,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 198,
      "max_score": null,
      "measured_at": "2026-07-13T10:55:25Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 221,
      "rating_method": "sequential_elo",
      "raw_elo": 967.8087609312524,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 5.347373388052779,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 23
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-05",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 183,
      "max_score": null,
      "measured_at": "2026-07-19T22:54:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 25,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 900.7652140057838,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 0.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 9
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 226,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 34,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1887.0029000547952,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 81.38419307216813,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 158
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 180,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 13,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 35,
      "max_score": null,
      "measured_at": "2026-07-13T11:13:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1811.769918810733,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 75.05378217838995,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 132
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 634,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 270,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 247,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 562,
      "rating_method": "sequential_elo",
      "raw_elo": 1570.038337646147,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.19569968343782,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 313
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 190,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 5,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 91,
      "max_score": null,
      "measured_at": "2026-07-18T13:28:06Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 190,
      "rating_method": "sequential_elo",
      "raw_elo": 1556.48816333833,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 54.10125469479228,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 94
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 82,
      "max_score": null,
      "measured_at": "2026-07-18T13:24:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1553.295615896295,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 53.83214821376729,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 97
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 193,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 10,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 92,
      "max_score": null,
      "measured_at": "2026-07-18T13:17:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 193,
      "rating_method": "sequential_elo",
      "raw_elo": 1531.2418555500674,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 52.04121036523053,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 91
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 192,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 8,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 93,
      "max_score": null,
      "measured_at": "2026-07-18T13:15:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 1498.4404656743122,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 49.33453889017528,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 91
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 599,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 291,
      "draws": 25,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 264,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 566,
      "rating_method": "sequential_elo",
      "raw_elo": 1494.4022754125206,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 47.55645299565777,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 19.379453129541623,
      "unilateral_faults": 29,
      "wins": 277
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 182,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 90,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 110,
      "max_score": null,
      "measured_at": "2026-07-18T13:22:41Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 182,
      "rating_method": "sequential_elo",
      "raw_elo": 1370.5796066184876,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 38.74378697017428,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 72
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 566,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 268,
      "draws": 3,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 334,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 542,
      "rating_method": "sequential_elo",
      "raw_elo": 1358.066580931903,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 37.704641705021,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 205
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 471,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 213,
      "draws": 9,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 287,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 418,
      "rating_method": "sequential_elo",
      "raw_elo": 1306.1824899049018,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.50727998813733,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 12.453570188460423,
      "unilateral_faults": 53,
      "wins": 122
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 121,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:23Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1287.166138448974,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 31.83539904740799,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 53
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 180,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 129,
      "max_score": null,
      "measured_at": "2026-07-13T10:58:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1236.0895422486844,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.657743183844648,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 49
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 4,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 127,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 1204.7661324053583,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.051367389220953,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 45
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 180,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 152,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1107.073403600676,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.0357831740134,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 28
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 180,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 14,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 136,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:54Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1081.256169780604,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.910238043962352,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 30
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 180,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 141,
      "max_score": null,
      "measured_at": "2026-07-13T11:18:03Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1077.5229142525063,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.602877342103772,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 38
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 180,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 149,
      "max_score": null,
      "measured_at": "2026-07-13T11:17:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1072.0730196388874,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 14.154184928501236,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 30
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 2,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 152,
      "max_score": null,
      "measured_at": "2026-07-19T22:38:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1046.3479252422885,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 12.036225680973985,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 26
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 180,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 89,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 147,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 180,
      "rating_method": "sequential_elo",
      "raw_elo": 1035.3016086951216,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.126777189882008,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 32
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 386,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 179,
      "draws": 1,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 309,
      "max_score": null,
      "measured_at": "2026-07-19T23:04:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 362,
      "rating_method": "sequential_elo",
      "raw_elo": 1034.1366678134975,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.03757348862904,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 52
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-06",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 168,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 176,
      "rating_method": "sequential_elo",
      "raw_elo": 900.4879497141699,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 0.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 215,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-20T08:39:29Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1847.9939468086293,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 74.21628479825006,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 142
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 198,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 47,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:11Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1788.9399080122687,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 70.14174273877752,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 151
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 620,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 282,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 214,
      "max_score": null,
      "measured_at": "2026-07-20T08:44:44Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 591,
      "rating_method": "sequential_elo",
      "raw_elo": 1653.2249464492245,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.61574859377792,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 377
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 199,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 67,
      "max_score": null,
      "measured_at": "2026-07-20T08:44:37Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 199,
      "rating_method": "sequential_elo",
      "raw_elo": 1648.799998835468,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 60.31532930836284,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 132
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 202,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 79,
      "max_score": null,
      "measured_at": "2026-07-20T08:52:02Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 202,
      "rating_method": "sequential_elo",
      "raw_elo": 1640.3837262661373,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 59.73960964866883,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 123
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 218,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 83,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:29Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 194,
      "rating_method": "sequential_elo",
      "raw_elo": 1596.1815838243494,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 56.59865223036674,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 111
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 210,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 94,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:41Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 210,
      "rating_method": "sequential_elo",
      "raw_elo": 1586.3464106719218,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.98636710066346,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 116
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 209,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 89,
      "max_score": null,
      "measured_at": "2026-07-18T14:36:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5-Pro",
      "model_slug": "xiaomi-mimo-v2-5-pro",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 209,
      "rating_method": "sequential_elo",
      "raw_elo": 1574.857203822943,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.17575767805614,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 120
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 88,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 98,
      "max_score": null,
      "measured_at": "2026-07-12T07:19:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1458.5825351906751,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 46.891914278195024,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 4,
      "wins": 86
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 223,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 113,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:41Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 199,
      "rating_method": "sequential_elo",
      "raw_elo": 1456.3249209298287,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 46.81240856410825,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 86
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 201,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 116,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:59Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 201,
      "rating_method": "sequential_elo",
      "raw_elo": 1427.7559875446782,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 44.81802752654815,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 85
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 208,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 117,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:09Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1365.238196858895,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 40.343423865130546,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 67
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 198,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 140,
      "max_score": null,
      "measured_at": "2026-07-13T10:08:42Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1295.044843928131,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 35.492966497854844,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 58
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 619,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 277,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 396,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 571,
      "rating_method": "sequential_elo",
      "raw_elo": 1273.7950614480271,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 33.96211011336977,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 175
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 209,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 142,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:24Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 185,
      "rating_method": "sequential_elo",
      "raw_elo": 1250.712260924394,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 32.31451053091617,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 43
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 398,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 183,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 288,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:48Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 374,
      "rating_method": "sequential_elo",
      "raw_elo": 1194.3117812051498,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 28.368736329667165,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 86
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 199,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 157,
      "max_score": null,
      "measured_at": "2026-07-20T08:44:35Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 199,
      "rating_method": "sequential_elo",
      "raw_elo": 1192.6283326358723,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 28.313005051674057,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 42
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 397,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 185,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 284,
      "max_score": null,
      "measured_at": "2026-07-20T08:38:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 397,
      "rating_method": "sequential_elo",
      "raw_elo": 1180.4988339222468,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 27.459581015327913,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 113
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 412,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 185,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 300,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 383,
      "rating_method": "sequential_elo",
      "raw_elo": 1159.1235825951264,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.92335810041868,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 83
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 442,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 198,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 344,
      "max_score": null,
      "measured_at": "2026-07-20T08:50:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 389,
      "rating_method": "sequential_elo",
      "raw_elo": 1139.5075927819269,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 22.935719462016753,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 25.83927083938883,
      "unilateral_faults": 53,
      "wins": 45
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 184,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 163,
      "max_score": null,
      "measured_at": "2026-07-20T08:51:32Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 184,
      "rating_method": "sequential_elo",
      "raw_elo": 1010.131975174144,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.431257472598286,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 21
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 198,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 174,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:38Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Plus",
      "model_slug": "qwen-qwen3-7-plus",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 198,
      "rating_method": "sequential_elo",
      "raw_elo": 1003.598767190432,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-plus",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 15.046821894435537,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 24
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 211,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 92,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 175,
      "max_score": null,
      "measured_at": "2026-07-20T08:43:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 187,
      "rating_method": "sequential_elo",
      "raw_elo": 957.13759060968,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.730010145194916,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 199,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 91,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-07",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 195,
      "max_score": null,
      "measured_at": "2026-07-20T08:43:51Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemini 3.1 Flash Lite",
      "model_slug": "google-gemini-3-1-flash-lite",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 24,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 199,
      "rating_method": "sequential_elo",
      "raw_elo": 789.0454370328994,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemini-3.1-flash-lite",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 0.0,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 0,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 0,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 21,
      "max_score": null,
      "measured_at": "2026-07-13T11:07:15Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.8",
      "model_slug": "claude-opus-4-8",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 1,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1963.6270727845756,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-8",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 80.76792349183744,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 168
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 19,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 19,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 15,
      "max_score": null,
      "measured_at": "2026-07-13T11:06:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Opus 4.5",
      "model_slug": "claude-opus-4-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 2,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1946.403774774844,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-opus-4-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 78.90896540288884,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 154
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 593,
      "capped_draws": 140,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 285,
      "draws": 140,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 102,
      "max_score": null,
      "measured_at": "2026-07-19T22:47:52Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GLM-5.2",
      "model_slug": "z-ai-glm-5-2",
      "plugin_success_rate": null,
      "provider": "Zhipu",
      "provider_key": "z-ai",
      "rank": 3,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 567,
      "rating_method": "sequential_elo",
      "raw_elo": 1730.5062551182962,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "z-ai",
        "provider_model": "z-ai/glm-5.2",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 55.71803693884606,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 325
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 195,
      "capped_draws": 25,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 25,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 58,
      "max_score": null,
      "measured_at": "2026-07-18T13:59:28Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Minimax M3",
      "model_slug": "minimax-minimax-m3",
      "plugin_success_rate": null,
      "provider": "MiniMax",
      "provider_key": "minimax",
      "rank": 4,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 195,
      "rating_method": "sequential_elo",
      "raw_elo": 1714.4927144474516,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "minimax",
        "provider_model": "minimax/minimax-m3",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 54.04543152948865,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 112
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 220,
      "capped_draws": 52,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 98,
      "draws": 52,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 47,
      "max_score": null,
      "measured_at": "2026-07-19T22:51:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Sol",
      "model_slug": "gpt-5-6-sol",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 5,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 196,
      "rating_method": "sequential_elo",
      "raw_elo": 1583.3259326638351,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-sol",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 39.95878166044322,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 97
    },
    {
      "admitted_plugins": 4,
      "attempted_matches": 822,
      "capped_draws": 402,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 4,
      "distinct_rated_opponents": 377,
      "draws": 402,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 177,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Nano",
      "model_slug": "gpt-5-4-nano",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 6,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 750,
      "rating_method": "sequential_elo",
      "raw_elo": 1488.4194514279093,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-nano",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 29.69233519321472,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 72,
      "wins": 171
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 200,
      "capped_draws": 124,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 101,
      "draws": 124,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 45,
      "max_score": null,
      "measured_at": "2026-07-18T13:36:57Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT 4.1",
      "model_slug": "gpt-4-1",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 7,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 200,
      "rating_method": "sequential_elo",
      "raw_elo": 1476.4580599067292,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-4.1",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 28.50597915073807,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 31
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 192,
      "capped_draws": 152,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 152,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 23,
      "max_score": null,
      "measured_at": "2026-07-18T13:22:40Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Pro",
      "model_slug": "deepseek-deepseek-v4-pro",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 8,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 192,
      "rating_method": "sequential_elo",
      "raw_elo": 1449.8521929305557,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 25.584865632593804,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 17
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 212,
      "capped_draws": 140,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 140,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 40,
      "max_score": null,
      "measured_at": "2026-07-19T23:05:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.5",
      "model_slug": "gpt-5-5",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 9,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1430.026546465291,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 23.42194383985197,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 214,
      "capped_draws": 126,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 126,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 44,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4",
      "model_slug": "gpt-5-4",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 10,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 190,
      "rating_method": "sequential_elo",
      "raw_elo": 1406.06609545599,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.86369919301597,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 20
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 212,
      "capped_draws": 128,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 128,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Terra",
      "model_slug": "gpt-5-6-terra",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 11,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1405.3189584689703,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-terra",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.76700401531612,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 17
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 218,
      "capped_draws": 146,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 97,
      "draws": 146,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 43,
      "max_score": null,
      "measured_at": "2026-07-19T22:55:56Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nex N2 Pro",
      "model_slug": "nex-agi-nex-n2-pro",
      "plugin_success_rate": null,
      "provider": "Nex AGI",
      "provider_key": "nex-agi",
      "rank": 12,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 191,
      "rating_method": "sequential_elo",
      "raw_elo": 1401.9036805195983,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nex-agi",
        "provider_model": "nex-agi/nex-n2-pro",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 20.42453892869399,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 2
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 613,
      "capped_draws": 360,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 284,
      "draws": 360,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 157,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:31Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.4 Mini",
      "model_slug": "gpt-5-4-mini",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 13,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 565,
      "rating_method": "sequential_elo",
      "raw_elo": 1394.4251675149944,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.4-mini",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 19.59912055423234,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 48,
      "wins": 48
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 189,
      "capped_draws": 129,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 129,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-13T11:03:10Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Nemotron 3 Ultra 550B A55B",
      "model_slug": "nvidia-nemotron-3-ultra-550b-a55b",
      "plugin_success_rate": null,
      "provider": "NVIDIA",
      "provider_key": "nvidia",
      "rank": 14,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1387.9420259336252,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "nvidia",
        "provider_model": "nvidia/nemotron-3-ultra-550b-a55b",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.90801399755653,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 12
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 216,
      "capped_draws": 132,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 96,
      "draws": 132,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 49,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:19Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "North Mini Code",
      "model_slug": "cohere-north-mini-code",
      "plugin_success_rate": null,
      "provider": "Other",
      "provider_key": "other",
      "rank": 15,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1386.7927190408395,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "other",
        "provider_model": "cohere/north-mini-code",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 18.78451587864971,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 24,
      "wins": 8
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 127,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 127,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 48,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:21Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Qwen3.7 Max",
      "model_slug": "qwen-qwen3-7-max",
      "plugin_success_rate": null,
      "provider": "Alibaba",
      "provider_key": "qwen",
      "rank": 16,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1371.829805973906,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "qwen",
        "provider_model": "qwen/qwen3.7-max",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 17.151771576470264,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 11
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 126,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 126,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 47,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:14Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Hy3 Preview",
      "model_slug": "tencent-hy3-preview",
      "plugin_success_rate": null,
      "provider": "Tencent",
      "provider_key": "tencent",
      "rank": 17,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1370.0230082937594,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "tencent",
        "provider_model": "tencent/hy3-preview",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.957623158720452,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 13
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 186,
      "capped_draws": 129,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 93,
      "draws": 129,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 49,
      "max_score": null,
      "measured_at": "2026-07-13T11:15:22Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Gemma 4 31B",
      "model_slug": "google-gemma-4-31b-it",
      "plugin_success_rate": null,
      "provider": "Google",
      "provider_key": "google",
      "rank": 18,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 186,
      "rating_method": "sequential_elo",
      "raw_elo": 1363.7606083498813,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "google",
        "provider_model": "google/gemma-4-31b-it",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.28470053225246,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 8
    },
    {
      "admitted_plugins": 3,
      "attempted_matches": 426,
      "capped_draws": 240,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 3,
      "distinct_rated_opponents": 209,
      "draws": 240,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 131,
      "max_score": null,
      "measured_at": "2026-07-19T23:12:07Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "DeepSeek V4 Flash",
      "model_slug": "deepseek-deepseek-v4-flash",
      "plugin_success_rate": null,
      "provider": "DeepSeek",
      "provider_key": "deepseek",
      "rank": 19,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 396,
      "rating_method": "sequential_elo",
      "raw_elo": 1381.1916633843164,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "deepseek",
        "provider_model": "deepseek/deepseek-v4-flash",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 16.090109959157065,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 21.786315039332067,
      "unilateral_faults": 30,
      "wins": 25
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 125,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 94,
      "draws": 125,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 57,
      "max_score": null,
      "measured_at": "2026-07-13T10:52:12Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "GPT-5.6 Luna",
      "model_slug": "gpt-5-6-luna",
      "plugin_success_rate": null,
      "provider": "OpenAI",
      "provider_key": "openai",
      "rank": 20,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1341.9363597433005,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "openai",
        "provider_model": "gpt-5.6-luna",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 13.95626289544574,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 6
    },
    {
      "admitted_plugins": 2,
      "attempted_matches": 375,
      "capped_draws": 239,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 2,
      "distinct_rated_opponents": 189,
      "draws": 239,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 132,
      "max_score": null,
      "measured_at": "2026-07-13T11:14:55Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "MiMo-V2.5",
      "model_slug": "xiaomi-mimo-v2-5",
      "plugin_success_rate": null,
      "provider": "Xiaomi",
      "provider_key": "xiaomi",
      "rank": 21,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 375,
      "rating_method": "sequential_elo",
      "raw_elo": 1321.438027318165,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "xiaomi",
        "provider_model": "xiaomi/mimo-v2.5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.749474217977998,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 4
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 190,
      "capped_draws": 130,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 130,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 56,
      "max_score": null,
      "measured_at": "2026-07-13T11:04:16Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Claude Sonnet 5",
      "model_slug": "claude-sonnet-5",
      "plugin_success_rate": null,
      "provider": "Anthropic",
      "provider_key": "anthropic",
      "rank": 22,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 189,
      "rating_method": "sequential_elo",
      "raw_elo": 1319.8723042635704,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "anthropic",
        "provider_model": "claude-sonnet-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.593620745823015,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 3
    },
    {
      "admitted_plugins": 1,
      "attempted_matches": 188,
      "capped_draws": 118,
      "codegen_cost_usd": null,
      "codegen_evaluated_cells": null,
      "conditional_success_score": null,
      "contributing_entrants": 1,
      "distinct_rated_opponents": 95,
      "draws": 118,
      "entry_count": null,
      "failed_game_aliases": [],
      "first_pass_admitted": null,
      "first_pass_rate": null,
      "first_seen_at": null,
      "game_alias": "GB2-08",
      "generation_policy": {
        "generation_policy": "quality_generation_policy",
        "stream_idle_timeout_seconds": null,
        "streaming_enabled": null,
        "strict_wall_clock_timeout_seconds": null,
        "timeout_mode": "provider_route_timeout_guard"
      },
      "generation_timing": {
        "finish_reason": "unknown_legacy",
        "first_progress_after_seconds": null,
        "generation_completed_at": null,
        "generation_started_at": null,
        "generation_wall_clock_seconds": null,
        "stream_progress_event_count": null
      },
      "infra_excluded_game_aliases": [],
      "infra_failures_excluded": null,
      "is_recent_addition": false,
      "last_measured_at": null,
      "losses": 64,
      "max_score": null,
      "measured_at": "2026-07-13T11:16:04Z",
      "min_score": null,
      "model_code_failure_rate": null,
      "model_code_failures": null,
      "model_display_name": "Mistral Medium 3.5",
      "model_slug": "mistralai-mistral-medium-3-5",
      "plugin_success_rate": null,
      "provider": "Mistral",
      "provider_key": "mistral",
      "rank": 23,
      "rank_movement": null,
      "rank_movement_baseline_at": null,
      "ranking_status": "not_applicable",
      "rated_matches": 188,
      "rating_method": "sequential_elo",
      "raw_elo": 1316.934684773728,
      "reasoning_level": "None",
      "reasoning_provenance": {
        "code_output_mode": "unknown_legacy",
        "code_prompt_version": "unknown_legacy",
        "generation_settings_public": {
          "max_tokens": null,
          "stream": null,
          "temperature": null
        },
        "prompt_identity": "unknown_legacy",
        "prompt_profile": "unknown_legacy",
        "provider": "mistral",
        "provider_model": "mistralai/mistral-medium-3-5",
        "provider_reasoning_params": {
          "requested_effort": "none",
          "source": "public_redacted_snapshot",
          "track_id": "full_freedom/none"
        },
        "provider_route": "unknown_legacy",
        "reasoning_bucket": "none",
        "reasoning_bucket_label": "No requested reasoning",
        "reasoning_policy_version": "gb2_reasoning_provenance_v1",
        "requested_reasoning_effort": "none"
      },
      "repaired_admitted": null,
      "repaired_rate": null,
      "results_cut_id": "gb2-20260720-959ae35a724b",
      "row_kind": "game",
      "score": 11.269722122355963,
      "score_basis": "relative per-game score",
      "scoring_policy_version": "entrant_identity_equal_mean_v7",
      "snapshot_id": "20260720_093741",
      "suite_line_id": "gamebench2",
      "suite_version": "2-known8-r2",
      "track_id": "full_freedom/none",
      "track_name": "None variants",
      "uncertainty": 0.0,
      "unilateral_faults": 0,
      "wins": 6
    }
  ],
  "schema_version": 7,
  "snapshot_id": "20260720_093741",
  "suite_line_id": "gamebench2",
  "suite_version": "2-known8-r2"
}