[Shuvam1024]

Summary

Post-hoc attribution for the SciFact chunked-BM25 gap, plus a timing refresh of the committed rerank reports. Frozen selections, thresholds, and the existing headline test numbers are unchanged.

• Document-level BM25 with the selected parameters (k1 1.5, b 0.75, stem) is scored on the SciFact test split and on NFCorpus. These runs were done after the frozen test numbers were reported. They are not a train selection and they do not replace the document-level baseline or the selected chunked index.
• On SciFact, that ablation's nDCG@10 is 0.6910, against 0.6741 for document-level BM25 at k1 0.9, b 0.4 (mean delta +0.017, 95% CI [+0.005, +0.028], sign-flip p 0.0045) and 0.6907 for the selected chunked index. The chunked index differs on 5 of 300 queries (mean delta -0.0003, 95% CI [-0.0007, -0.00004], p 0.0621). The +0.017 gap is the parameter change. Chunking adds about nothing on this split.
• NFCorpus gets the same ablation. The parameter-change interval includes zero. Chunking versus the matched parameters also leaves the nDCG interval covering zero (12 of 323 queries differ).
• benchmarks/scifact/rerank-selected.json and benchmarks/scifact/rerank-train.json were regenerated with the current index clock. Cross-encoder weight loading is outside index_ms. Ranking metrics, including nDCG@10 and the rerank delta of -0.028, match the previous reports. Per-query latencies moved with the new wall clock, and the rendered p95 figures follow them.

Test plan

• ruff check and ruff format --check
• mypy
• pytest with coverage (89.09% with the dense extra installed locally)
• README and docs/results.md drift checks
• CI jobs on this pull request