# Validation record — 2026-09-30

1. Frozen PROTOCOL.md registered before NFCorpus download/inference. Existing082–086 sources match priorZIP; no duplicate087 directory.
2. Initial network request exit7 (sandbox proxy unreachable); network-authorized retry exit0, officialMD5+SHA256 pinned. BEIR/models reused and hash checked. acquisition.json records download time unmeasured.
3. First execution omitted--smoke: performed full BM25, then failed at query completeness assertion before dense inference or aggregate metrics. Preserved initial_failed and initial_failed.log. Upstream BM25 omits query keys without hits. Fixed only empty query retention and zero metric denominator. Direct ES same multi_match count confirmed all15 absent queries have0hits. No score-driven tuning; target debugging exposure explicitly disclosed.
4. Correct smoke5queries/32docs, exit0. Neural inference .3754163s; rough scaled forecast40.139s; formal ceiling900s/<4GiB PythonRSS. resource_decision.json records actual forecast; no performance conclusion from reduced-corpus scores (allzero relevance here).
5. Formal323queries/full3633corpus exit0: encoding37.1345105s,total42.6543916s,RSS1044201472B. Three methods,12334graded positive qrels. 18official fixed source reads/commit checks passed. All46852RRF union scores and969query-method metric rows audited, max3.3306690738754696e-16. Shape [2,8]→[2,8,384]→[2,384], masked-mean maxerror0. Query/document vectors[323,384]/[3633,384]. 2864documents truncated to256tokens;0queries truncated. Truncation is unchanged configuration, not evidence-loss annotation.
6. Formal execution identity captured modules before audit.py and documentation were added; run.py/encoder.py/fusion.py/compat.py/common.py unchanged after formal run. prepare.py later gained optional existing082 ES downloader; prediction path unchanged. Package manifest captures final source. No claim that audit.py generated the original inference.
7. Raw input data stays in explicit reader cache; archive/public copy includes every required self-authored module. Fresh-extraction syntax/help/offline audit/real smoke and remote file hashes recorded outside code in verification/ and publication_verification.json, avoiding circular hash identities.
8. No human labels, training, reranker run, confidence interval, dynamic leaderboard or SciFact300confirmation inference. Source809development scores reused, not unseen data. New targettest now consumed for transfer/debugging; any future selection on it is exploratory.