{
  "scope": "calibration_only",
  "cases": 400,
  "casesPerDisposition": 100,
  "holdoutUsed": false,
  "configurationFingerprint": "162acbcd233e665f45502e9f883c547faef4fd96246665e5dcb0b6891f35117e",
  "sourceCommit": "7e886ea8f9e6c0960961746adaf83dabbb02aa5c",
  "sources": [
    {
      "file": "final_calibration_evidence.json",
      "sha256": "b1884ccebdf2ed6a8b39191299041ec27b4e247e984aeca0f607048b34a2e0ec"
    },
    {
      "file": "inhouse_calibration_e2e_evaluation.json",
      "sha256": "c8fa1f5e54ea0594b4b3f2e00f44b8d1e7fa3bffabff02988d3fc281fdd63343"
    },
    {
      "file": "inhouse_judge_validation_v23.json",
      "sha256": "9326e278c7d651f978176083467e8c95702e72a5a08e8678fe72cf1e39aee90e"
    }
  ],
  "techniques": [
    {
      "id": "baseline",
      "name": "Baseline RAG",
      "correct": 100,
      "macroF1": 0.1,
      "answerRecall": 1.0,
      "rowInterval": [0.2075, 0.29],
      "familyInterval": [0.1107, 0.5152]
    },
    {
      "id": "regex_only_with_shared_controls",
      "name": "Regex scenario",
      "correct": 141,
      "macroF1": 0.271,
      "answerRecall": 1.0,
      "rowInterval": [0.305, 0.3975],
      "familyInterval": [0.1678, 0.6312]
    },
    {
      "id": "fuzzy_only_with_shared_controls",
      "name": "Fuzzy scenario",
      "correct": 153,
      "macroF1": 0.333,
      "answerRecall": 1.0,
      "rowInterval": [0.3375, 0.43],
      "familyInterval": [0.2294, 0.5935]
    },
    {
      "id": "bge_similarity_with_shared_controls",
      "name": "Semantic scenario",
      "correct": 249,
      "macroF1": 0.583,
      "answerRecall": 0.87,
      "rowInterval": [0.5725, 0.67],
      "familyInterval": [0.3431, 0.8339]
    },
    {
      "id": "deterministic_hybrid",
      "name": "Deterministic hybrid",
      "correct": 250,
      "macroF1": 0.584,
      "answerRecall": 0.87,
      "rowInterval": [0.575, 0.6725],
      "familyInterval": [0.3431, 0.8386]
    },
    {
      "id": "qwen_classifier_only",
      "name": "Classifier scenario",
      "correct": 389,
      "macroF1": 0.972,
      "answerRecall": 0.89,
      "rowInterval": [0.955, 0.9875],
      "familyInterval": [0.9614, 0.99],
      "meanLatencyMs": 23417.48
    },
    {
      "id": "complete_inhouse_hybrid",
      "name": "Complete hybrid",
      "correct": 392,
      "macroF1": 0.98,
      "answerRecall": 0.92,
      "rowInterval": [0.965, 0.9925],
      "familyInterval": [0.9667, 0.9964],
      "meanLatencyMs": 22502.85
    }
  ],
  "failures": {
    "retrieval": 2,
    "answerability": 2,
    "entailment": 4
  },
  "citations": {
    "total": 303,
    "expectedDocumentPrecision": 0.766,
    "verifierConditionedPrecision": 1
  },
  "judge": {
    "cases": 200,
    "exactAgreement": 0.795,
    "dimensions": {
      "grounded": 0.915,
      "privacy_safe": 0.995,
      "injection_safe": 0.995,
      "integrity_safe": 0.9,
      "refusal_appropriate": 0.97
    },
    "humanAnnotators": 1,
    "assistedPasses": 2
  }
}