[Dong1017]

Scope

Evidence-only follow-up to #255, with a consolidated efficiency/accuracy table and the three previously unpublished studies. No production source, scorer, corpus or default behavior changes; no new experiments were run for this packaging. Previously published #256/#259 results remain unchanged. This is evidence delivery, not accepted retrieval improvement; #255 remains open.

Study	Preserved result
Structured-string token v3	FAIL: Hit@5 4/6→5/6, MRR7/12→8/15; development gate failed, heldout not evaluated
Blind raw-task/cue preparation	INCONCLUSIVE: only3/6 source-supported pairs, below frozen4-case minimum; no retrieval execution
Source-conditioned covered cue	Neutral FAIL: both arms Hit6/6 and MRR29/36; strictMRR-gain gate failed

The unified table also links historical timingv1FAIL, separate timingconfirmation with wide uncertainty, and accuracyv2FAIL. It does not pool or relabel those outcomes. Source-conditioned cases and reused trainingDEV are explicitly distinguished from independent validation. The sealed unused heldout content is absent.

Artifacts and provenance

• Frozen protocols, inputs, exact raw gzip outputs, runner/source receipts and independent audits.
• Rejected v3 runtime retained only as a patch/Git bundle, not installed into production.
• Two blind-preparation local metadata fields replaced by labelled placeholders; original/public hashes and unchanged task text documented.
• Historical source snapshots retain exact bytes with documented replay filename mapping.
• Three classified evidence bundles add79 hashes; inventory totals311 hashes/14bundles.
• Main table: consolidated results.

Verification

• Bounded evidence inventory tests:17passed.
• Ruff on affected test/validator files:passed; git diff --check:passed.
  -79newhashes,28gzip JSON captures, consolidated report/protocol references and navigation links verified.
• Independent packaging audit passed at1468e628; finalc3aa9540 changes only replay-path clarification plus its hash, with affected tests passing again. HostedCI not run yet.

No timing/model/service/hardware experiments, spending, runtime promotion, formal approval or Issue acceptance is implied.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-30T05:16:22.799828Z	00d6f0e	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[Dong1017]

Scoped LGTM for evidence delivery at 00d6f0e (base 6e6ee24). The remote tree matches the independently audited package: 79 added manifest hashes, 28 gzip JSON captures, preserved v3 FAIL / blind-preparation INCONCLUSIVE / covered-cue neutral FAIL, and separate historical timing outcomes. No production source changes or sealed heldout contents. The sole packaging suggestion (consistent restored PREP_ROOT instructions) was adopted and rechecked; current native Codex review 6515560061 reports no unresolved findings, and no other review threads remain.

Verification: 17 bounded evidence tests plus integrity/reference checks passed; Ready CI run 41007369198 passed full-check and installed-package jobs on Python 3.10/3.11, and knowledge-model-policy passed. Earlier cancelled draft CI is superseded. The PR-body hosted-CI statement describes submission time; these are the subsequent results.

Limits: this accepts archival/packaging correctness, not retrieval improvement, generalization, agent efficacy or campaign acceptance. No new experiments were run during packaging. #255 remains open; efficiency and accuracy goals remain unfulfilled. COMMENT only, not a formal APPROVE.