← Back
📝 Notes / reports / phase3a-analogue-transfer-20260919

📑 Contents

Phase 3A structure-aware analogue transfer with mass-only fallback — 2026-09-19ScopeResultsInterpretationRequired audit before promotion

Phase 3A structure-aware analogue transfer with mass-only fallback — 2026-09-19


Scope


Bounded post-freeze stability diagnostic over 1,171 deterministic identity-holdout molecules and 474 valid queries. This is not prospective validation, an official score, or a trained model.


The primary route uses eligible cross-identity spectral neighbours and scores PubChem candidates with:


max cosine(query, reference) × Morgan-Tanimoto(candidate structure, reference structure)


For queries with no eligible neighbour, the diagnostic now uses a deployable mass-only fallback: candidates are ordered by absolute neutral-mass deviation and deterministic identity14 tie-break. The fallback does not use target-relative fields or target identity for scoring.


Results


PolicyRankable spectral queriesNo spectral neighbourSpectral Recall@25 on rankableFallback Recall@25Combined Recall@25Combined MRR
P0 strict ±10 ppm3819386.61%89.25%87.13%0.3440
P1 Na ±30 ppm3819386.61%89.25%87.13%0.3440
P2 empty-pool fallback3819386.61%89.25%87.13%0.3440

For the 93 no-neighbour queries, the mass-only fallback recovered the target in the candidate pool for 90.32% and in the top 25 for 89.25%; fallback MRR was 0.7349. These are conditional fallback diagnostics, not evidence that the target was spectrally ranked.


Interpretation


  • Adding the mass-only fallback substantially reduces the unavailable-route gap in this stability sample.
  • The combined result is still dominated by the post-freeze cohort and should not be promoted as prospective performance.
  • P0/P1/P2 remain identical for this structure-aware route. Sodium-specific widening provides no additional measured benefit here.
  • The result supports a two-stage non-neural baseline: structure-aware analogue transfer when eligible neighbours exist, then mass-only ordering when they do not.

    • Required audit before promotion


  • Add exact failure-category counts: structure absent, no eligible spectral neighbour, fallback-ranked, and spectral-ranked.
  • Verify fallback candidate ordering against the independent deterministic mass-ranking implementation.
  • Add formula-gated fallback as a separate ablation; do not use oracle target formula at deployment.
  • Repeat on a genuinely pre-registered untouched identity partition.
  • Run maintained tests and add regression coverage for no-neighbour fallback and target-relative-field exclusion.

    • Artifact: reports/phase3a-analogue-transfer-20260919.json.

    Script: scripts/diagnose_phase3a_analogue_transfer.py.