{
  "status": "passed",
  "selected_method": "four_difference",
  "seeds": [
    2026,
    2027,
    2028,
    2029,
    2030
  ],
  "primary_runs": 400,
  "same_checkpoint_controls": 160,
  "verified_curve_points": 118160,
  "main_pairs": 16,
  "verified_confidence_curve_points": 3232,
  "sample_standard_deviation_ddof": 1,
  "metric_and_call_tolerance": 1e-12,
  "verified_pair_objective_summary_rows": 80,
  "raw_logit_score_check_absolute_tolerance": 8e-07,
  "counts_use_strict_improvement_above_tolerance": {
    "higher_metric_than_confidence": 13,
    "fewer_calls_than_confidence": 11,
    "higher_metric_and_fewer_calls": 10,
    "above_both_endpoints": 8
  },
  "selection": "validation task-balanced Score25 mean minus sample SD, before test replay",
  "scope": "Saved router outputs, frozen endpoint outcomes, explicit per-event correctness labels and specialist confidence/validity. Both learned and confidence curves and validation-selected points are independently replayed; no model retraining.",
  "study_history": "User changed selection criterion after the seed-42 pilot and partial CUB test results had been reported. The final numerical criterion uses validation records only; this is not a prospectively untouched test set."
}