← Back

📑 Contents

Analog-propagation review — decisions and plan revision1. Evidence inspected2. Revised next-work order3. Concrete notebook lessons mapped to TODOs4. Corrections to earlier conversational conclusions5. Local issues that must not be hidden by the new roadmap6. Files changed and verification boundary

Analog-propagation review — decisions and plan revision


Date: 22 September 2026 (MYT).

Request: review [prvsiyan's CASMI 2026 notebook](https://www.kaggle.com/code/prvsiyan/analog-propagation-casmi-2026-baseline), identify useful improvements, and update EXECUTION-PLAN.md and TODO.md.

Outcome: revised research priorities and falsifiable acceptance tasks. No new model training, heavy inference, notebook execution/push, dataset expansion, production promotion or automation changes.


1. Evidence inspected


  • [Notebook technical review](notebook-review.md): code/narrative review with cell and source-line anchors.
  • [Local evidence audit](local-evidence-review.md): current source, small saved manifests/results, NPY headers and synthetic ablation counterexamples. Source inspection and synthetic tests are not a full training/evaluation rerun.
  • [Source retrieval manifest](../../research/reviews/analog-propagation-20260922-1902/retrieval-manifest.json).
  • [Downloaded notebook](../../research/reviews/analog-propagation-20260922-1902/raw/analog-propagation-casmi-2026-baseline.ipynb) and [metadata](../../research/reviews/analog-propagation-20260922-1902/raw/kernel-metadata.json).
  • Original plan/TODO copies and their hashes are preserved in [before/](before/).
  • Saved official rules in research/pages.json, especially the competition-specific winner-license and external-data/tool clauses; this was not a fresh legal approval or a new rules scrape.

  • The notebook was retrieved at 19:04:12–14 MYT and matched the previously retained 15:15 source exactly. SHA-256: 3ce550ab18441573ee519a4e1df88ae03090e00e424c16671e02bc6b76214a9e. The pull did not establish a version number. It contains 37 cells, zero saved outputs, and null execution counts. All performance tables are author-reported; none were reproduced here. Six attached dataset names do not prove asset licensing, availability, training provenance or fold isolation.


    2. Revised next-work order


    1. Repair the comparison contract: scorer-identity grouping, exposure ledger, baseline effective configuration, ablation semantics and independent audit checks. Preserve useful historical evidence without overclaiming gate closure.

    2. Freeze a bounded eligible candidate pool and non-neural controls: use a documented natural-product-focused source plus permitted training structures; independently measure membership coverage. Do not resume unbounded PubChem slice acquisition.

    3. Prioritize spectrum-to-fingerprint learning: binary fingerprint BCE, candidate score f · z, then same-mass and same-formula hard-negative ranking loss. Compare compatible multi-spectrum and domain-balanced variants.

    4. Add mass-shifted analogue evidence in parallel: transfer structural evidence from eligible chemical neighbours onto candidate structures. This is not exact-library lookup and not copying a neighbour's SMILES.

    5. Calibrate complementary channels: OOF ranker against single-channel/static controls; same-formula diagnostics; bounded fragment explanation and candidate–peak cross-encoding only if justified by errors and latency.

    6. Qualify, then consider extending the generator: validate the current objective and output contract first. Predicted-fingerprint/formula/valence conditioning is a possible follow-on, not an automatic long training run or a fixed five-slot fallback.


    The user has approved a documentation review, not execution of every new experiment. Previously authorized Phase 4 work is not stopped, relaunched or promoted by this revision. Model training remains off the local ARM host, including smoke training.


    3. Concrete notebook lessons mapped to TODOs


    Lesson / code evidenceAdopted workAcceptance / caution
    Four complementary channels; wider mass-shifted analogue search, C006/C007/C012FP-01/02/03 and AN-01/02Same folds/pool; identity-wide reference exclusions; retain per-channel evidence. Current narrow direct-cosine transfer is not already this route.
    Multi-family binary fingerprints and BCE + 63 same-mass decoys, C009/C010/C001:TRAINFP-01/02Compare Morgan control and documented bit union. f · z is independent-Bernoulli log-likelihood up to a query-only constant, not calibrated exact correctness. Audit false-negative/duplicate sampling.
    Reference formula-derived mass option, C007/C020CHEM-02Compare eligible reference metadata to measured precursor mass; keep unavailable query truth formulas out of inference.
    Multi-spectrum benefits but unsafe mixed-adduct metadata, C018FP-03Preserve each surviving spectrum's metadata; compatible ion groups or explicit set models, matched training/inference distributions.
    Provenance-feature leakage and attached ranker features, C022/C026RIGHTS-01, FUSE-01No in-sample upstream features or dataset-origin shortcuts; retain licensing/OOF provenance and group-aware calibration.
    Unused knobs and cap changes affecting all channels, C002/C012/C026/C027BASE-02, FUSE-02Bind effective parameters to execution, record answer retention and regenerate/refit set-relative features after pool/pruning changes.
    Failure-conditioned isomer benchmark is biased, C033ISO-01, RERANK-01Define full/cohort/isomer panels independently of ranker failures; report shortlist retention before reranker quality.
    Unsuccessful logits–fingerprint interaction heads, C032RERANK-01A candidate–peak cross-encoder is a distinct hypothesis; use OOF trunk features and compare against simple scoring under latency constraints.
    DreaMS not an automatic analogue upgrade, C031EMB-01Optional challenger only; the reported negative test changed analogue similarity, not fingerprint-head fine-tuning.
    Rerun score differences and unknown test mixture, C025/C035REPRO-01, VAL-02Estimate our own numerical/seed uncertainty; no universal leaderboard noise threshold or inferred exact class labels.

    4. Corrections to earlier conversational conclusions


  • Class 1 versus 2: public reference-spectrum availability differs from whether we have an eligible local reference. Class 2 concerns known structures without public reference spectra; it is not necessarily a newly invented structure. The estimated 0.151 / 0.93 ≈ 16% public Class-1 share rests on transfer assumptions. The remainder can include both Class 2 and Class 3; it is not a known 84% Class-2 fraction.
  • Coverage: the notebook's COCONUT 99.6% figure concerns a particular 250-compound sample. A benchmark with answers retained in its pool cannot demonstrate hidden-test coverage. Increasing a small CID slice does not mechanically double the probability of covering a test compound.
  • Generation: exact-connectivity scoring does not require an identical canonical SMILES spelling and does not make generation intrinsically wrong. A generator can identify in-pool or out-of-pool molecules. Beam width 25 is not 25 valid unique identities and does not directly optimize MRR. Use calibrated union/reranking, not a mandated 20/5 split.
  • DreaMS: C031's n=130 analogue comparison reports entropy 0.1555, DreaMS cosine 0.1129, and entropy×sqrt(DreaMS) 0.1428. It is not a failed fingerprint-fine-tuning experiment or proof of universal failure. The claimed mass-shift-invariance explanation is a hypothesis.
  • Reported performance: the notebook's fingerprint/analogue/combined values (~0.517/~0.52/~0.63) are author-reported simulations. They cannot be compared directly with our ~0.87 library-ceiling diagnostic, and cannot be described as our demonstrated gains.
  • Commercial use: paper, code, weights and datasets have distinct rights. Copyleft is not automatically non-commercial; permissive code does not automatically license linked weights/training data. Do not assume a competition qualifies for every academic-use exemption or that a published algorithm is unpatented.

  • 5. Local issues that must not be hidden by the new roadmap


    The local audit found substantive contract issues; it did not alter source or repair them in this review:


    1. Disputed baseline lock: configured 20 ppm versus 30 ppm evidence; 200 declared peaks versus 64 in the index; raw split bound versus scorer split consumed; existing entry-point defaults still at 10 ppm. All ten listed hashes match, demonstrating that hash consistency alone is insufficient.

    2. Ablation defects reproduced on synthetic records: widest-window identity aggregation can delete valid narrow-window evidence; a binding mass cap leaves mass order as final rank; target rank 1001 is incorrectly counted as not in the pool. Candidate counts also mean unique identities rather than spectral comparisons. Existing mass/cap/compute conclusions need corrected replay, not a new label.

    3. Exposure/namespace: saved G1 records contain 1,163 scorer-fold-0 identities in each library/B cohort and 47 in A; Phase 4 repeatedly validates raw-key fold 0. These facts contradict a blanket untouched-fold claim, without proving every fold-0 identity leaked into training.

    4. Generator source versus acceptance: shifted targets and unconditional causal masking are present locally. Raw-identity grouping, greedy top-1 metrics, optional-RDKit fallback and missing immutable run binding remain acceptance issues. The earlier claimed RDKit auto-install is not in either current local source copy.

    5. Audit scope: stronger exact cohort, mask and row-identity joins and fail-on-drift behavior are needed for the independent G1 acceptance audit. That deficiency alone is not evidence the historical run leaked or drifted.

    6. Do not redo fixed work: all-spectrum terminal-checkpoint ordering is now corrected in source. Genuine resume state remains unfinished; the old before-metrics ordering bug is not still an implementation TODO.


    These are tracked as VAL-01/02, BASE-01/02, AUDIT-01, GEN-01/02 and PKG-01. Existing artifacts remain intact and unpromoted. This task has not changed the remote kernel, its current status or its monitoring schedule.


    6. Files changed and verification boundary


    Updated:


  • [EXECUTION-PLAN.md](../../EXECUTION-PLAN.md): current priorities, class/coverage caveats, fingerprint/analogue objectives, chemistry/metadata handling, feature and validation contracts, generator role, rights review and revised execution ladder. Original calendar and early data inspection are explicitly historical.
  • [TODO.md](../../TODO.md): concrete task IDs, dependencies, P0 acceptance repairs, source-present versus validated status, bounded-pool preference, and corrected stale blocker/progress language. New experiments remain unchecked.

  • Supporting evidence consists of this report, the two reviews, the read-only notebook snapshot/metadata and original-document backups. The [independent final QA](final-qa.md) requested two scope/staleness corrections; both were applied: G1-05 now claims only the exclusion check its auditor actually performed, and G1-08/A6/precursor/packaging descriptions distinguish historical blockers from later completed fixes. Its two minor wording corrections were also applied.


    Final [verification.json](verification.json) records document/report hashes, link/task checks and QA-fix assertions; [changes.diff](changes.diff) preserves the final document diff. All 26 new experiment/acceptance tasks remain pending. Verification here is documentation/static evidence verification, not a full model test suite or G1/G4 promotion.