{
  "baselines": {
    "officialRerankNdcg": 0.8268,
    "endToEndNdcgAt10": 0.2597,
    "endToEndJudgedCoverageAt10": 0.2724,
    "publishedOfficialBaseline": 0.8503,
    "publishedKddWinner": 0.9043,
    "note": "Official rerank mode is directly comparable to published cross-encoder numbers. End-to-end nDCG is depressed by judgment incompleteness (only 27% of top-10 carry an ESCI label at full corpus scale); judged-coverage is reported so the depression is visible - it is a floor, not a quality measurement."
  },
  "method": "BM25 multi_match best_fields over title^3/brand^2/color/bullets/description, tie_breaker 0.3; official ESCI graded gains E/S/C/I; judged-coverage@k reported for end-to-end.",
  "artifacts": [
    "experiments/esci/baseline-bm25-gen032.json",
    "experiments/esci/baseline-bm25-full-gen033.json"
  ]
}