{
  "schema_version": 1,
  "name": "Brali Bench",
  "bench_version": "1.1.0",
  "dataset_version": "1.1.0",
  "case_count": 50,
  "scope": "Deterministic evaluation of Brali retrieval, grounding, provenance, evidence boundaries, and conservative no-answer behavior.",
  "not_a_claim": "This is not a benchmark of an unpinned language model and does not measure general model intelligence.",
  "comparison_layers": [
    {
      "id": "no-knowledge-control",
      "purpose": "Control showing what is available without Brali knowledge."
    },
    {
      "id": "lexical-brali",
      "purpose": "Keyword-oriented Brali retrieval baseline."
    },
    {
      "id": "structured-brali",
      "purpose": "Ontology-, protocol-, and Evidence-Decision-aware Brali retrieval."
    }
  ],
  "required_gates": [
    "safety and deliberate no-answer behavior",
    "trusted evidence-state preservation",
    "reviewed-source provenance preservation",
    "zero unsupported Evidence Decision claims",
    "structured retrieval usefulness must not regress below the lexical baseline"
  ],
  "reproduce": {
    "build": "npm run build",
    "verify": "npm run evaluate:check",
    "runner": "scripts/run-agent-evaluation.mjs",
    "checker": "scripts/check-agent-evaluation.mjs"
  },
  "artifacts": {
    "cases": "cases.json",
    "results": "results.json",
    "methodology": "https://github.com/Brali-LifeOS/brali-lifeos.github.io/blob/main/docs/AGENT_EVALUATION.md",
    "canonical_page": "https://brali-lifeos.github.io/bench/"
  },
  "license": "CC-BY-NC-SA-4.0",
  "citation": "Dzmitryi Kharlanau. Brali Practical Knowledge Library, pinned data-v release."
}