[Shuvam1024]

Summary

Interview-readiness corrections from an independent audit. Committed benchmark metric values, train selections, and thresholds are unchanged except where a step below adds an optional field or changes which frozen config the CI gate runs.

• Coverage numbers are no longer committed. The README points at the CI badge and the 85% coverage gate. benchmarks/pytest-junit.xml, benchmarks/test-coverage.json, and the unused requirements/constraints-py312.txt are removed.
• SciFact dev verdicts keep the existing per-claim sentence F1 and also report micro sentence F1 and mean sentence F1 over claims that have gold evidence. Both-empty claims still score 1.0 in the per-claim mean; that caveat is stated next to the number. A dev-only oracle feeds gold evidence documents to the frozen NLI policy, and an error breakdown separates retrieval misses from NLI errors. Nothing was selected on dev.
• The chunk-tuning claim is restated from the data: 46 of 5,183 SciFact documents (26 of 3,633 NFCorpus) exceed the 480-word window, so the selected index is nearly one chunk per document. The test gap versus document-level BM25 is mostly the BM25 parameters (1.50/0.75 vs 0.90/0.40). Those counts are rendered from the reports.
• README is a short entry point. Methodology, selection tables, NFCorpus, exploratory runs, verdict details, and the judge sample live in docs/results.md. Table regions in both files are rendered and drift-checked. Acronyms are defined on first use. Ragas, DeepEval, promptfoo, and ARES have citation links.
• CI gates the frozen document-level BM25 config and the frozen selected BM25 config, still with a zero point-drop policy. The exploratory configs/scifact-bm25.yaml is no longer the gate. Hybrid stays out of the main job because it needs the dense extra.
• pytrec_eval is an optional agreement extra. The agreement test skips without it. The Linux CI job installs g++ and that extra. Install instructions are uv only.
• Evaluation reports are assembled in one path. Newly written comparison files can store the statistics recipe; committed comparison files were not rewritten.
• Offline tests inject fake encoders for dense search, embeddings, NLI, and rerank. ragstat.__version__ comes from package metadata. Rerank index_ms excludes model loading. CITATION.cff is added.

Test plan

• ruff check, ruff format --check, mypy
• pytest with coverage, gated extras only (no dense, no agreement)
• README figure drift test across README.md and docs/results.md
• Support-fixture evaluate and compare against benchmarks/baseline.json (shared non-timing fields)
• SciFact document-level BM25 and selected BM25 compare, zero point-drop
• docker build and docker run
• uv sync --frozen --extra dev --extra api without building pytrec_eval
• CI jobs on this pull request