{
  "protocol": "Frozen P=100 geometric shortlist -> top-10 RAG ranking (HQ10 soft branch diversity lambda_D=0.1 where applicable).",
  "datasets": [
    "SciFact",
    "TREC-COVID",
    "Quora",
    "MS MARCO / DL19",
    "HotpotQA",
    "NQ"
  ],
  "OURS_CPU_100_to_10": {
    "nDCG@10": [
      0.5685,
      0.599,
      0.7366,
      0.34,
      0.467,
      0.2579
    ],
    "MRR@10": [
      0.5452,
      0.8252,
      0.7287,
      0.5189,
      0.6255,
      0.2262
    ],
    "P@10": [
      0.0737,
      0.652,
      0.1124,
      0.2674,
      0.0964,
      0.0462
    ],
    "R@10": [
      0.6663,
      0.0163,
      0.8407,
      0.0916,
      0.482,
      0.3983
    ],
    "Hit@10": [
      0.6833,
      1.0,
      0.8904,
      0.7209,
      0.7507,
      0.4342
    ],
    "median_ms": [
      0.947,
      1.051,
      120.301,
      65.195,
      28.944,
      33.686
    ],
    "p95_ms": [
      1.031,
      1.193,
      175.494,
      142.647,
      42.568,
      44.161
    ]
  },
  "notes": [
    "The six-dataset benchmark row is the frozen 100->10 operating point.",
    "The postbenchmark final_two_optimizations branch is preserved separately because it changes some NQ/Hotpot values and is not the source of the frozen table.",
    "MS MARCO / DL19 is the 43-query TREC-DL19 run over the full 8.84M-passage corpus."
  ]
}