{
  "title": "RAG Production Acceptance Benchmark 2026: Which Release Gates Catch Unsupported Answers?",
  "dataset": "ParticleMedia/RAGTruth commit c103204b9ce28d6bbad859304bf30de72b8ed8fe",
  "datasetUrl": "https://github.com/ParticleMedia/RAGTruth",
  "license": "MIT",
  "evaluatorVersion": "pran-rag-gates/1.0.0",
  "executionDate": "2026-08-23",
  "protocol": "Official RAGTruth train split calibrates each threshold; official test split is reported. Only records with quality=good are included. Labels marked implicit_true are not counted as hallucinations.",
  "trainResponses": 14942,
  "testResponses": 2675,
  "positiveTestResponses": 894,
  "methods": [
    {
      "id": "corpus_coverage",
      "name": "Corpus token coverage",
      "description": "Blocks a response when its overall unsupported content-token share exceeds a train-calibrated threshold.",
      "threshold": 0.34,
      "precision": 0.5278,
      "recall": 0.8065,
      "f1": 0.6381,
      "falseNegativeRate": 0.1935,
      "tp": 721,
      "fp": 645,
      "tn": 1136,
      "fn": 173,
      "meanFeatureLatencyMs": 0.0229,
      "measuredProviderCostUsd": 0
    },
    {
      "id": "weakest_sentence",
      "name": "Weakest-sentence support",
      "description": "Blocks a response when any sentence of five or more content tokens falls below the train-calibrated support threshold.",
      "threshold": 0.58,
      "precision": 0.5179,
      "recall": 0.8255,
      "f1": 0.6365,
      "falseNegativeRate": 0.1745,
      "tp": 738,
      "fp": 687,
      "tn": 1094,
      "fn": 156,
      "meanFeatureLatencyMs": 0.0229,
      "measuredProviderCostUsd": 0
    },
    {
      "id": "hybrid_claim_gate",
      "name": "Hybrid claim gate",
      "description": "Combines overall support, weakest-sentence support, and source-missing numeric claims in a fixed weighted score.",
      "threshold": 0.44,
      "precision": 0.5371,
      "recall": 0.8188,
      "f1": 0.6486,
      "falseNegativeRate": 0.1812,
      "tp": 732,
      "fp": 631,
      "tn": 1150,
      "fn": 162,
      "meanFeatureLatencyMs": 0.0229,
      "measuredProviderCostUsd": 0
    }
  ],
  "limitation": "These are intentionally simple local lexical gates, not a claim that lexical overlap establishes truth. They cannot reliably detect contradictions, paraphrases, or unsupported claims built from source vocabulary."
}