{
  "baseline": {
    "metrics": {
      "accuracy": 0.5061728395061729,
      "average_precision": 0.5368354184411922,
      "balanced_accuracy": 0.5047011193141224,
      "brier_score": 0.3265273012310616,
      "count": 648,
      "expected_calibration_error": 0.2555114427420742,
      "f1": 0.05325443786982249,
      "false_negative_rate": 0.9721362229102167,
      "false_positive_rate": 0.018461538461538463,
      "fn": 314,
      "fp": 6,
      "log_loss": 1.0058404875183675,
      "macro_f1": 0.35961260515620563,
      "negative_count": 325,
      "positive_count": 323,
      "precision": 0.6,
      "recall": 0.02786377708978328,
      "roc_auc": 0.5538747320790665,
      "specificity": 0.9815384615384616,
      "threshold": 0.991553083857187,
      "tn": 319,
      "tp": 9
    },
    "name": "acoustic_logistic_source_holdout",
    "predictions": {
      "bytes": 140627,
      "path": "artifacts/partial-source-holdout-baseline/validation_predictions.jsonl",
      "sha256": "aa459b9221c30b25dc747959a5963ee588ddbd7e69dbad2d740a039fbbbf6913"
    }
  },
  "bootstrap": {
    "samples": 2000,
    "seed": 17,
    "unit": "audit leakage group"
  },
  "candidate": {
    "metrics": {
      "accuracy": 0.5123456790123457,
      "average_precision": 0.5499581204674366,
      "balanced_accuracy": 0.5108930697785187,
      "brier_score": 0.28202905638850445,
      "count": 648,
      "expected_calibration_error": 0.16490815351084795,
      "f1": 0.07602339181286549,
      "false_negative_rate": 0.9597523219814241,
      "false_positive_rate": 0.018461538461538463,
      "fn": 310,
      "fp": 6,
      "log_loss": 0.7782068838177573,
      "macro_f1": 0.37239324726911616,
      "negative_count": 325,
      "positive_count": 323,
      "precision": 0.6842105263157895,
      "recall": 0.04024767801857585,
      "roc_auc": 0.5522743510359609,
      "specificity": 0.9815384615384616,
      "threshold": 0.7510702013969421,
      "tn": 319,
      "tp": 13
    },
    "name": "tiny_tcn_source_holdout",
    "predictions": {
      "bytes": 210723,
      "path": "artifacts/partial-source-holdout-tinytcn/validation_predictions.jsonl",
      "sha256": "f19fa98ab11be1ceeda0d49100972797aaa6e8747beb84bab91a614e2625b6f4"
    }
  },
  "candidate_minus_baseline_95ci": {
    "average_precision": {
      "bootstrap_fraction_delta_above_zero": 0.574,
      "estimate": 0.013122702026244437,
      "lower": -0.21491768842361025,
      "upper": 0.12612715782503575
    },
    "false_positive_rate": {
      "bootstrap_fraction_delta_above_zero": 0.393,
      "estimate": 0.0,
      "lower": -0.012671805155199063,
      "upper": 0.025158553987632778
    },
    "recall": {
      "bootstrap_fraction_delta_above_zero": 0.7695,
      "estimate": 0.01238390092879257,
      "lower": -0.011235955056179775,
      "upper": 0.09483031427324311
    },
    "roc_auc": {
      "bootstrap_fraction_delta_above_zero": 0.5025,
      "estimate": -0.0016003810431055587,
      "lower": -0.2712768963129158,
      "upper": 0.12730811695916328
    }
  },
  "count": 648,
  "fpr_budget": 0.02,
  "group_count": 9,
  "negative_count": 325,
  "positive_count": 323,
  "scope": "paired development comparison; not independent-test evidence",
  "threshold_note": "Each threshold was selected on this same development set. Bootstrap samples use those fixed thresholds; intervals do not remove model-selection or calibration bias."
}