{
  "verdict": "V10A_AUDIT_CORRECTED_LABELS_CONFIRMS_V10_POSITIVE",
  "audit_sha256": "04606e1ae9828fe77ec66c30d967829609e523bb23a6bdfdfc0e170e9c87a30d",
  "original_v10_protocol_sha256": "0ea7c2ac6252e4206ab2099033056b8dd2f977f1984d861ca3a0f6c4e1848fc0",
  "original_v10_verdict": "V10_THREE_REGIME_PROSPECTIVE_TWO_DOOR_BEATS_EARLY_STOPPING",
  "frozen_v09b_policy_sha256": "9b0b8ab801d2c6b52efdf38835960240dbf27d6a057f3a14dc0f9af3328c64af",
  "n_runs": 9,
  "label_changes": {
    "count": 1,
    "run_ids": [
      "T01"
    ],
    "details": [
      {
        "run_id": "T01",
        "original": "STABLE_BENEFICIAL",
        "corrected": "NO_ENTRY",
        "best_raw_validation_step": 0,
        "policy_decision_step": 60
      }
    ]
  },
  "corrected_regime_counts": {
    "stable_beneficial": 3,
    "exit_after_entry": 3,
    "no_entry": 3
  },
  "safety_detection": {
    "actions_on_intended_beneficial": 0,
    "exit_detected": "3/3",
    "no_entry_detected": "3/3",
    "prebest_actions": 0,
    "exact_policy_restores": "6/6",
    "median_rollback_test_regret_pct": 0.0
  },
  "compute_saving": {
    "macro_policy_pct": 43.827160493827165,
    "macro_es_pct": 35.18518518518519,
    "macro_advantage_points": 8.641975308641975,
    "weighted_policy_pct": 50.0,
    "weighted_es_pct": 41.666666666666664,
    "weighted_advantage_points": 8.333333333333336
  },
  "robustness_excluding_label_disagreement_case": {
    "excluded_run_ids": [
      "T01"
    ],
    "macro_policy_pct": 40.97222222222223,
    "macro_es_pct": 34.02777777777778,
    "macro_advantage_points": 6.94444444444445,
    "weighted_policy_pct": 48.484848484848484,
    "weighted_es_pct": 41.41414141414141,
    "weighted_advantage_points": 7.070707070707073
  },
  "criteria": {
    "integrity_checks_pass": true,
    "at_least_2_stable_beneficial_runs": true,
    "at_least_2_exit_after_entry_runs": true,
    "at_least_2_no_entry_runs": true,
    "zero_actions_on_intended_beneficial": true,
    "exit_sensitivity_100pct": true,
    "no_entry_sensitivity_100pct": true,
    "no_prebest_actions": true,
    "median_rollback_test_regret_le_1pct": true,
    "all_policy_restores_exact": true,
    "macro_saving_beats_es": true,
    "step_weighted_saving_beats_es": true,
    "positive_without_disagreement_case_macro": true,
    "positive_without_disagreement_case_weighted": true
  },
  "scope": "V10a changes no policy parameter and performs no retraining. It corrects one derived edge label and confirms that the positive V10 compute and safety conclusions remain unchanged."
}