← Back
📝 Notes / reports / plan-review-20260922-analog / final-qa

📑 Contents

Final documentation QA — 22 September 2026Required fixesMinor clarity caveatsOtherwise satisfactory

Final documentation QA — 22 September 2026


Result: two bounded documentation fixes required.


Scope: read all five requested documents (EXECUTION-PLAN.md, TODO.md, REVIEW.md, notebook-review.md, local-evidence-review.md). This is an internal-consistency check against the supplied reviews, not an independent rerun of their source/data findings. No network, model work, real-data scan, code changes, commits or automation changes. Only this QA report was written.


Required fixes


1. Narrow the G1-05 exclusion claim to what its audit actually checked. TODO.md:108 reports Regime B as “identity-wide exclusion, 0 violations”. However, local-evidence-review.md P0-10 says the post-run auditor only checks candidate != current target, not exclusion of every held-out fold identity, and does not independently join rows to identity/masks. AUDIT-01 correctly leaves that stronger audit pending. Preserve the completed arithmetic audit and its numerical results, but replace the parenthetical with wording such as “0 current-target candidate matches under the existing limited audit; full fold/mask exclusion audit pending AUDIT-01”. Do not imply the stronger audit already passed; equally, do not infer actual leakage from its missing checks.


2. Reconcile residual present-tense implementation blockers with the recorded completed fixes. TODO.md:53 still leaves precursor coercion open, and :112 presents malformed-precursor handling/packaging as remaining blockers, although G1-04 at :107 records the null-precursor fix and expanded local packaging acceptance and G1-07 records a later offline smoke. Explicitly date G1-08's blocker list as the 19 September decision, cross-reference its later dispositions, and distinguish any additional untested precursor cases/final packaging requirements from the fix already present. Similarly, G1-02 at :105 still says “Promote these probes into maintained tests (A6)” while A6 at :65 is completed: point to the completed coverage and leave only demonstrably missing cases pending. This does not close full G1 or G6.


Minor clarity caveats


  • EXECUTION-PLAN.md:51–53 calls the retained leaderboard/account figures a “Live snapshot” and says “At review time” beneath a 22 September revision date. Label this explicitly as the original 16 September snapshot, rather than implying a refreshed leaderboard/account check.
  • EXECUTION-PLAN.md:316 says “Add precursor/adduct/energy…” although the local review confirms existing precursor-m/z input. Prefer “extend existing precursor conditioning with adduct/energy and predicted formula…”; do not imply the whole input path is absent.

    • Otherwise satisfactory


  • Author-reported notebook scores are not presented as reproduced gains or compared directly with our library-ceiling diagnostic.
  • Raw and scorer-canonical identities, reference availability and structure-pool coverage, and the previously exposed folds are distinguished.
  • Baseline config/peak-width/split/default conflicts and ablation ordering/coverage defects remain pending repairs, not new baseline certification.
  • Shifted generator targets, unconditional causal masking and terminal-checkpoint ordering are recognized as already present; behavioral qualification and genuine resume state remain open.
  • No local training, automatic bulk catalogue expansion, fixed generator slot quota, blanket commercial-use permission, or full G1/G4/G6 closure is authorized or claimed.
  • Newly proposed fingerprint, analogue, fusion, pool, rights, generator and packaging experiments remain unchecked. Existing Phase 4 work is retained conditionally without asserting a fresh remote status.