{
  "schema": "flamehaven_api_v1_bav_metrics",
  "generated_at_utc": "2026-06-28T09:52:58.779666+00:00",
  "note": "SR9 >= 0.70 pipeline guard, >= 0.80 honesty gate. DI2 <= 0.30 pipeline guard, <= 0.20 honesty gate. p_e2e = capture x transfer x model x clinical.",
  "experiments": [
    {
      "id": "bav-exp-034",
      "title": "Cross-cycle non-degradation check",
      "verdict": "PASS",
      "date": "2026-04-19",
      "cross_domain_consistency": 1.0,
      "reasoning_deviation": 0.127,
      "end_to_end_reliability": 0.6072,
      "balanced_accuracy": 1.0,
      "pass_recall": 1.0,
      "block_recall": 1.0
    },
    {
      "id": "bav-exp-033",
      "title": "Repro2 parity test against baseline",
      "verdict": "FAIL",
      "date": "2026-03-08",
      "cross_domain_consistency": 0.9531,
      "reasoning_deviation": 0.1637,
      "end_to_end_reliability": 0.5627,
      "balanced_accuracy": 0.5,
      "pass_recall": 0.0,
      "block_recall": 1.0
    },
    {
      "id": "bav-exp-032",
      "title": "PASS/BLOCK classification parity",
      "verdict": "PASS",
      "date": "2026-04-16",
      "balanced_accuracy": 1.0,
      "dangerous_pass_rate": 0.0,
      "pass_recall": 1.0,
      "block_recall": 1.0
    },
    {
      "id": "bav-exp-031",
      "title": "Out-of-distribution protein target \u2014 3-arm observer study",
      "verdict": "HOLD",
      "date": "2026-02-09",
      "drift_mean": 0.1864,
      "plddt_mean": 52.17,
      "arms_unverified": 3,
      "arms_total": 3
    },
    {
      "id": "bav-exp-028",
      "title": "Honesty calibration under cross-domain query",
      "verdict": "ABSTAIN",
      "date": "2026-02-05",
      "cross_domain_consistency_auc": 1.0,
      "overall_auc": 1.0,
      "calibration_brier_score": 0.0056,
      "calibration_ece_score": 0.075,
      "cross_domain_consistency_threshold": 0.8,
      "cross_domain_consistency_gate_passed": false
    },
    {
      "id": "bav-exp-005",
      "title": "Upadacitinib topical formulation null screen",
      "verdict": "NULL",
      "date": "2025-12",
      "cross_domain_consistency_mean": 0.254,
      "cross_domain_consistency_threshold": 0.8,
      "samples_rejected": 3,
      "samples_total": 3
    }
  ]
}