{
  "_meta": {
    "version": "v1",
    "generated_at": "2026-07-22T12:52:19.913Z",
    "total": 3000,
    "categories": [
      "math",
      "date",
      "text",
      "entity",
      "refusal",
      "adversarial"
    ],
    "generator": "scripts/build-hallucination-benchmark.js",
    "seed": "0x5CE05"
  },
  "run_at": "2026-08-06T06:32:52.188Z",
  "url": "http://localhost:4444",
  "model": "cogos-tier-b",
  "duration_ms": 117,
  "total": 3001,
  "pass": 2993,
  "fail": 8,
  "pass_rate": 0.9973,
  "hallucination_rate": 0.0027,
  "ci_95": {
    "lo": 0.9947,
    "hi": 0.9986
  },
  "by_category": {
    "math": {
      "pass": 1105,
      "total": 1105,
      "primitive_hits": 1105,
      "pass_rate": 1,
      "hallucination_rate": 0,
      "ci_95": {
        "lo": 0.9965,
        "hi": 1
      },
      "deterministic_coverage_pct": 100
    },
    "date": {
      "pass": 600,
      "total": 600,
      "primitive_hits": 600,
      "pass_rate": 1,
      "hallucination_rate": 0,
      "ci_95": {
        "lo": 0.9936,
        "hi": 1
      },
      "deterministic_coverage_pct": 100
    },
    "text": {
      "pass": 249,
      "total": 249,
      "primitive_hits": 249,
      "pass_rate": 1,
      "hallucination_rate": 0,
      "ci_95": {
        "lo": 0.9848,
        "hi": 1
      },
      "deterministic_coverage_pct": 100
    },
    "entity": {
      "pass": 143,
      "total": 147,
      "primitive_hits": 11,
      "pass_rate": 0.9728,
      "hallucination_rate": 0.0272,
      "ci_95": {
        "lo": 0.9321,
        "hi": 0.9894
      },
      "deterministic_coverage_pct": 7.5
    },
    "refusal": {
      "pass": 397,
      "total": 400,
      "primitive_hits": 0,
      "pass_rate": 0.9925,
      "hallucination_rate": 0.0075,
      "ci_95": {
        "lo": 0.9782,
        "hi": 0.9974
      },
      "deterministic_coverage_pct": 0
    },
    "adversarial": {
      "pass": 499,
      "total": 500,
      "primitive_hits": 496,
      "pass_rate": 0.998,
      "hallucination_rate": 0.002,
      "ci_95": {
        "lo": 0.9888,
        "hi": 0.9996
      },
      "deterministic_coverage_pct": 99.2
    }
  },
  "cache": {
    "hits": 3001,
    "misses": 0,
    "hit_rate": 1,
    "no_cache_mode": false
  }
}