{
  "updated": "2026-10-12",
  "source_files": [
    {
      "file": "run-023/seed-1000/heldout.csv",
      "sha256": "495228da24661ad09a9958a49d3fc2e9beb3c44a2c4d6540f2e1d4e74ab87c16"
    },
    {
      "file": "run-023/seed-1000/metrics.json",
      "sha256": "f047cfa5865246f23eb39d256cf00510af9ab5928e8b60a1953aa1ffd144a3a2"
    }
  ],
  "rows": 3000,
  "positive_rows": 238,
  "method": "For each arm: numpy.quantile(test negative scores, 0.95), then strict score > threshold. Thresholds use test labels.",
  "nominal_false_positive_budget": 0.05,
  "results": [
    {
      "system": "Starting classifier",
      "score_column": "p_seed",
      "threshold": 0.15,
      "recall": 0.7310924369747899,
      "false_positive_rate": 0.017016654598117305
    },
    {
      "system": "Direct GEPA optimization",
      "score_column": "p_raw_cal",
      "threshold": 0.1348290191806124,
      "recall": 0.7016806722689075,
      "false_positive_rate": 0.023171614771904415
    },
    {
      "system": "Gnosys",
      "score_column": "p_wrap_cal",
      "threshold": 0.10181499059081293,
      "recall": 0.7773109243697479,
      "false_positive_rate": 0.0391020999275887
    }
  ],
  "recall_gain_vs_starting_classifier_percentage_points": 4.621848739495804,
  "precision_recall_auc": {
    "seed_ap": 0.7650628462383138,
    "raw_ap": 0.720559240504764,
    "wrap_ap": 0.6921848276497139
  },
  "limitations": [
    "Actual false-positive rates differ due to tied scores.",
    "Does not establish equal realized FPR, equal review cost, or a deployable threshold.",
    "Headline run only; no new training or experiment performed.",
    "Two early single-seed studies overall; not a multi-seed replication."
  ]
}
