{
  "updated_at": "2026-08-24",
  "protocol": {"dataset_rows": 1499, "primary_rows": 1198, "released_positive_label": "T3 ready-to-review only", "conceptual_positive_label": "T2 needs-polish plus T3 ready-to-review", "zero_shot": true, "cache_disabled": true},
  "completed_conditions": {
    "auto_router_a": {"accuracy": 0.532554, "precision": 0.455934, "recall": 0.8, "f1": 0.580838, "false_positive_rate": 0.649369, "cost_usd": 0.22945015876},
    "auto_router_b": {"accuracy": 0.538397, "precision": 0.461009, "recall": 0.828866, "f1": 0.592483, "false_positive_rate": 0.659187, "cost_usd": 0.23688033376},
    "ox_alpha": {"accuracy": 0.603506, "precision": 0.506443, "recall": 0.810309, "f1": 0.623315, "false_positive_rate": 0.537167, "recorded_cost_usd": 0.0},
    "inkling_small": {"accuracy": 0.616861, "precision": 0.525794, "recall": 0.546392, "f1": 0.535895, "false_positive_rate": 0.335203, "confusion_matrix": {"tp": 265, "fp": 239, "tn": 474, "fn": 220}, "cost_usd": 2.74779095, "input_tokens": 1728916, "output_tokens": 2015948}
  },
  "auto_router_repeatability": {"mean_accuracy": 0.535476, "accuracy_delta_percentage_points": 0.5843, "mean_false_positive_rate": 0.654278, "prediction_agreement": 0.880634, "cohen_kappa": 0.704636}
}