{
  "schema_version": "1.1.0",
  "status": "research_only",
  "denominator": {
    "cases": 7583,
    "authorities": 39,
    "metric": "balanced_accuracy",
    "caveat": "Reused development periods; not untouched final test."
  },
  "rows": [
    {
      "model": "weak_authority_past_only_expert_router",
      "balanced_accuracy_full": 0.6927434422666074,
      "balanced_accuracy_recent": 0.6043201689753414,
      "equal_authority_balanced_accuracy_full_min_5_each_class": 0.6602396628114757,
      "eligible_authorities": 37,
      "authorities_at_or_above_80_percent": 1,
      "role": "best_unpromoted_routed_candidate",
      "promotion_status": "research_only_probability_regression_recent_minimum_gain_full_ci_and_target_noninferiority_gates_failed_development_reused",
      "comparison_to_prior_best": {
        "balanced_accuracy_full_delta": 0.0022882434920435957,
        "balanced_accuracy_recent_delta": 0.0008841732979664174,
        "full_authority_year_bootstrap_ci": [
          -0.0006269442873534749,
          0.0055126275350812805
        ],
        "recent_authority_year_bootstrap_ci": [
          -0.006100671524058583,
          0.008323485290614038
        ],
        "prior_best_brier": 0.4399008726375279,
        "candidate_brier": 0.4421995441316846,
        "prior_best_log_loss": 0.7102188627762062,
        "candidate_log_loss": 0.7130316945179829,
        "full_target_authorities_improved_worsened_unchanged": [
          6,
          1,
          1
        ],
        "worsened_full_authority": "ISO:NO"
      },
      "evidence": {
        "report_path": "reports/weak_authority_ensemble_threshold_v8/REPORT.md",
        "report_sha256": "623a00bcbd195072f41151e6fc0756c9092aed0162c4407a58b0291b1080822a",
        "receipt_path": "reports/weak_authority_ensemble_threshold_v8/receipt.json",
        "receipt_sha256": "20fb9755473e66f8db6b203fa68119b8c71d5692fc39438b0427a93cdfbd985b"
      }
    },
    {
      "model": "all_authority_regularized_hard_best_expert",
      "balanced_accuracy_full": 0.692864413190469,
      "balanced_accuracy_recent": 0.5913891344925828,
      "equal_authority_balanced_accuracy_full_min_5_each_class": 0.6594825147712872,
      "eligible_authorities": 37,
      "role": "rejected_higher_full_only_candidate",
      "promotion_status": "rejected_recent_equal_authority_per_authority_and_confidence_gates_failed_development_reused",
      "comparison_to_current_best": {
        "balanced_accuracy_full_delta": 0.00012097092386154085,
        "balanced_accuracy_recent_delta": -0.012931034482758563,
        "full_authority_year_bootstrap_ci": [
          -0.0030994340209965134,
          0.0031142644877198257
        ],
        "recent_authority_year_bootstrap_ci": [
          -0.03374581152099179,
          0.004175895311116627
        ],
        "current_best_brier": 0.4421995441316846,
        "candidate_brier": 0.44182321500390553,
        "current_best_log_loss": 0.7130316945179829,
        "candidate_log_loss": 0.7122444691214165,
        "full_authorities_improved_worsened_unchanged": [
          9,
          11,
          19
        ]
      },
      "evidence": {
        "report_path": "reports/all_authority_regularized_expert_router_v8/REPORT.md",
        "report_sha256": "392db1b9ab3798ffd08968785778a633259ce71ff8633b544fee2fc9ff94fa5b",
        "receipt_path": "reports/all_authority_regularized_expert_router_v8/receipt.json",
        "receipt_sha256": "068aff9d5b8e8d657807705bd6349b8718e932b4853410e77acb5a65193c5783"
      }
    },
    {
      "model": "conditional_hurdle__current_router_conditional_benchmark__partial_hurdle_010",
      "balanced_accuracy_full": 0.6920284023230933,
      "balanced_accuracy_recent": 0.6101164161508988,
      "role": "rejected_conditional_hurdle_candidate",
      "promotion_status": "rejected_full_direction_and_class_recall_gates_failed_despite_probability_and_recent_improvement",
      "comparison_to_current_best": {
        "balanced_accuracy_full_delta": -0.0007150399435141841,
        "balanced_accuracy_recent_delta": 0.00579624717555749,
        "full_authority_year_bootstrap_ci": [
          -0.0044685875022171225,
          0.0027028184642471616
        ],
        "current_best_brier": 0.4421995441316846,
        "candidate_brier": 0.419165738139426,
        "current_best_log_loss": 0.7130316945179829,
        "candidate_log_loss": 0.6783207625736977,
        "candidate_class_recall_cut_hold_hike": [
          0.6269487750556793,
          0.7048039382108301,
          0.7443324937027708
        ]
      },
      "evidence": {
        "report_path": "reports/conditional_action_direction_v8/REPORT.md",
        "report_sha256": "fcf37068c784972449118310f47aac42b05706168f8e2383984a43c76730446c",
        "receipt_path": "reports/conditional_action_direction_v8/receipt.json",
        "receipt_sha256": "9945efc6c87f6b199bd5852b7e51f080d46bbf25239c211ecb204cf9be15f92b"
      }
    },
    {
      "model": "chronological_probability_stacker_nested_fallback",
      "balanced_accuracy_full": 0.6927434422666074,
      "balanced_accuracy_recent": 0.6043201689753414,
      "equal_authority_balanced_accuracy_full_min_5_each_class": 0.6602396628114757,
      "eligible_authorities": 37,
      "authorities_at_or_above_80_percent": 1,
      "role": "rejected_probability_stacker_exact_fallback",
      "promotion_status": "rejected_all_fixed_stacks_lower_direction_nested_selector_retained_current_router_all_years",
      "fixed_stack_tradeoff": {
        "best_direction_variant": "fixed_stack_c0010_a025",
        "best_direction_balanced_accuracy_full": 0.6885137251278252,
        "best_direction_brier": 0.41508519129626364,
        "best_direction_log_loss": 0.6734282078722313,
        "lowest_brier_variant": "fixed_stack_c0100_a100",
        "lowest_brier_balanced_accuracy_full": 0.6600666081875913,
        "lowest_brier": 0.35700475676646504,
        "lowest_brier_log_loss": 0.5909846209601433
      },
      "evidence": {
        "report_path": "reports/chronological_probability_stacker_v8/REPORT.md",
        "report_sha256": "199c2c6d45cbea64021fb343bb171d58251abf35220e0f779c5f76190cae4759",
        "receipt_path": "reports/chronological_probability_stacker_v8/receipt.json",
        "receipt_sha256": "3383ddedb068664db8cf39688aabd3223fdd7003de9d114d8d05a8cea650bc84"
      }
    },
    {
      "model": "policy_event_boj_history_prior_only_hard_router",
      "balanced_accuracy_full": 0.6904551987745638,
      "balanced_accuracy_recent": 0.6034359956773749,
      "equal_authority_balanced_accuracy_full_min_5_each_class": 0.6546457426877516,
      "eligible_authorities": 37,
      "authorities_at_or_above_80_percent": 1,
      "role": "prior_best_unpromoted_routed_candidate",
      "promotion_status": "positive_ci_vs_canonical_but_not_secure_vs_previous_best_and_reuses_development"
    },
    {
      "model": "policy_event_prior_only_hard_router",
      "balanced_accuracy_full": 0.68840458,
      "balanced_accuracy_recent": 0.59664751,
      "equal_authority_balanced_accuracy_full_min_5_each_class": 0.65355742,
      "eligible_authorities": 37,
      "authorities_at_or_above_80_percent": 1,
      "role": "prior_unpromoted_routed_candidate",
      "promotion_status": "failed_frozen_minimum_gain_and_confidence_gates"
    },
    {
      "model": "chronological_model_zoo_stack",
      "balanced_accuracy_full": 0.68792022,
      "balanced_accuracy_recent": 0.58886433,
      "equal_authority_balanced_accuracy_full_min_5_each_class": 0.651486,
      "eligible_authorities": 37,
      "authorities_at_or_above_80_percent": 0,
      "role": "unpromoted_ensemble_candidate",
      "promotion_status": "confidence_interval_crosses_zero"
    },
    {
      "model": "ppi_availability_extension",
      "balanced_accuracy_full": 0.68761241,
      "balanced_accuracy_recent": 0.58694862,
      "equal_authority_balanced_accuracy_full_min_5_each_class": 0.65104,
      "eligible_authorities": 37,
      "role": "research_champion",
      "promotion_status": "canonical_research_comparator"
    },
    {
      "model": "policy_event_full_features",
      "balanced_accuracy_full": 0.68350438,
      "balanced_accuracy_recent": 0.60344828,
      "role": "rejected_full_history_candidate",
      "promotion_status": "full_balanced_accuracy_decreased"
    },
    {
      "model": "original_strong_baseline",
      "balanced_accuracy_full": 0.683204,
      "role": "baseline"
    },
    {
      "model": "monthly_retraining",
      "balanced_accuracy_full": 0.681276,
      "role": "rejected_capacity_tweak"
    },
    {
      "model": "peer_rate_breadth",
      "balanced_accuracy_full": 0.67779,
      "role": "rejected"
    },
    {
      "model": "larger_trees",
      "balanced_accuracy_full": 0.676762,
      "role": "rejected"
    }
  ],
  "component_metrics": [
    {
      "component": "documented_policy_action_opportunity_occurrence",
      "model": "hist_gradient_boosting",
      "balanced_accuracy_full": 0.7756724237873103,
      "balanced_accuracy_recent": 0.8184796232291389,
      "metric_scope": "binary occurrence proxy; not cut/hold/hike direction accuracy",
      "target_caveat": "missing historical unchanged-rate meeting archives can create false-negative occurrence labels",
      "direction_recombination_status": "rejected; best full direction balanced accuracy 0.6871036264152776 versus 0.6904551987745638 router"
    },
    {
      "component": "conditional_direction_on_documented_opportunity_rows",
      "model": "conditional_hist_gradient_boosting",
      "balanced_accuracy_full": 0.6398587104016652,
      "balanced_accuracy_recent": 0.5914001015486164,
      "metric_scope": "cut/hold/hike direction within documented-opportunity-positive rows only",
      "target_caveat": "incomplete unchanged-meeting archives can omit true hold opportunities from conditional training",
      "benchmark_status": "below current-router conditional benchmark at 0.692938 full and 0.604498 recent"
    }
  ],
  "sources": [
    "reports/full_champion_macro_groups_v7/RESULTS.md",
    "reports/action_opportunity_occurrence_v8/REPORT.md",
    "reports/action_opportunity_occurrence_v8/receipt.json",
    "reports/weak_authority_ensemble_threshold_v8/REPORT.md",
    "reports/weak_authority_ensemble_threshold_v8/receipt.json",
    "reports/policy_event_boj_history_ablation_v8/receipt.json",
    "reports/policy_event_boj_history_routing_v8/receipt.json",
    "reports/policy_event_routing_v8/receipt.json",
    "reports/best_candidate_router_v8/receipt.json",
    "reports/chronological_model_zoo_v8/receipt.json",
    "reports/policy_event_ablation_v8/receipt.json",
    "config/research_champion_comparator.json",
    "reports/all_authority_regularized_expert_router_v8/REPORT.md",
    "reports/all_authority_regularized_expert_router_v8/receipt.json",
    "reports/conditional_action_direction_v8/REPORT.md",
    "reports/conditional_action_direction_v8/receipt.json",
    "reports/chronological_probability_stacker_v8/REPORT.md",
    "reports/chronological_probability_stacker_v8/receipt.json"
  ]
}