← Back
📝 Notes / reports / todo-review-20260918 / REVIEW

📑 Contents

TODO review — 2026-09-18DecisionsConfirmed current-code issuesEvidence and scope

TODO review — 2026-09-18


Reviewed EXECUTION-PLAN.md and reports/kaggle-community-review-findings-2026-09-17.md in full. Updated TODO.md, retaining existing completed tasks and historical experiment rows. Neither input document nor running evaluation code was changed.


Decisions


  • Correct and audit G1 before closure; execution exit is not gate acceptance.
  • Promote pinned structure-only snapshots, mass/formula candidates and non-neural fingerprint transfer ahead of learned ranking.
  • Add OOF fingerprint ranking and a separate mass-shifted analogue track; defer de novo and cross-route fusion.
  • Add placeholder, cross-library leakage, adduct-alias, cap/window, aggregation, provenance and offline-output checks.

    • Confirmed current-code issues


    probe-current-contract.py calls the current evaluator with mocked tiny arrays and no training/index data. Its result semantic-probes.json reproduces Regime-A query expansion/self-comparison, the conditional-MRR denominator error, full-pool coverage being truncated to top-1000 recall, and strict-output validation gaps. These are findings to fix, not passing G1 tests. Source hashes are attached to the probes.


    Evidence and scope


  • Original TODO: TODO.before-review.md
  • Consistency checks: review-check.json
  • Synthetic reproduction: probe-current-contract.py, semantic-probes.json
  • No production source edits, job interruption, model training, paid compute or submission in this review.
  • G1 and the existing goal remain open.