{
  "dataset": "MDACE ICD-10-CM",
  "split": "official test after benchmark preprocessing",
  "evaluation_unit": "individual note",
  "notes": 115,
  "gold_assignments": 634,
  "catalog": {
    "direct_program_codes": 898,
    "source": "union of labels across the processed dataset, including the test split",
    "per_note_test_labels_used_for_optimisation": false
  },
  "task_model": "Qwen/Qwen3.6-35B-A3B",
  "task_temperature": 0.6,
  "task_top_p": 0.95,
  "seed": 1,
  "optimisation": {
    "train_notes": 120,
    "validation_notes": 60,
    "trajectories_per_setting": 1,
    "matrix_runs": 18,
    "reflection_labels": [
      "Luna",
      "Terra",
      "Sol"
    ]
  },
  "scoring": {
    "code_normalisation": "trim whitespace, uppercase, deduplicate",
    "precision": "TP / (TP + FP)",
    "recall": "TP / (TP + FN)",
    "micro_f1": "2 * TP / (2 * TP + FP + FN)",
    "failures": "retain failed executions as empty predictions; all gold codes on those notes count as false negatives",
    "ranges": "scores are fractions in [0, 1]"
  },
  "provenance": {
    "verification": "aggregate counts recomputed from saved predictions; no new model inference",
    "matched_test_identifiers_sha256": "9184516ea18f8df81423a94cba7ea25ef6f2b30261c496c620a5065413bfe1a9",
    "source_system_summary_sha256": "6d4c11cc96ce2e42b0007683061037fc6a2aeb209aa1137d8138d1a04b600129",
    "prediction_checksums": "each results.csv row identifies its source prediction artifact; original CLH and faithful DSPy files also contain other splits, which were excluded from test scoring"
  },
  "privacy": "Aggregate measurements and checksums only. No clinical text, patient or note identifiers, row-level labels, model responses, demonstrations, or credentials are included.",
  "limitations": [
    "Dataset-derived restricted catalogue; not a complete-ontology evaluation.",
    "One optimisation trajectory and one final test pass per setting; no optimiser-variance estimate.",
    "Selected-program results; the best row is descriptive, not an independent test of a prespecified winner.",
    "Hospital-note results do not establish GP, NHS, or clinical deployment performance."
  ],
  "csv_rows": 21
}
