← Back
📝 Notes / reports / phase3a-analogue-fallback-audit-20260919

📑 Contents

Analogue fallback audit — 2026-09-19ScopeFailure categoriesFormula ablationDecision

Analogue fallback audit — 2026-09-19


Scope


Independent audit of the deterministic mass-only fallback and an oracle formula-gating ablation on the post-freeze stability sample. The audit selected 475 valid queries (the source row selection differs by one query from the 474-query analogue-transfer run because this audit reads molecular_formula alongside the other fields). This is not prospective validation.


Failure categories


P0 strict ±10 ppm:


  • valid queries: 475
  • target present in structure pool: 447
  • target absent from structure pool: 28
  • queries with eligible cross-identity neighbour: 260
  • fallback/no eligible-neighbour route: 215
  • fallback target recovered in pool: 194 / 215 = 90.23%
  • fallback Recall@25: 191 / 215 = 88.84%
  • fallback rank P50/P95: 1 / 7.35
  • fallback ordering equivalence mismatches: 0

    • The fallback therefore agrees exactly with the independent mass-order implementation on this audit sample.


    Formula ablation


    Oracle formula gating produced exactly the same result as mass-only fallback:


  • oracle formula pool recall: 90.23%
  • oracle formula Recall@25: 88.84%

    • This is not a deployable formula result. The query formula was read from visible training metadata, and the runtime query contract currently provides only adduct and precursor m/z for candidate generation. The equality suggests that, on this sample, formula gating does not improve the target’s mass-ordered position; it does not justify removing formula inference from the future plan.


    Decision


  • Keep mass-only fallback as the current deterministic fallback diagnostic.
  • Do not promote the formula ablation: it is oracle-only and showed no gain here.
  • Preserve the 28 structure-pool misses, 215 no-neighbour cases, and 260 spectral-route cases as explicit failure classes.
  • Repeat the complete route on a genuinely pre-registered untouched identity partition before baseline promotion.

    • Artifact: reports/phase3a-analogue-fallback-audit-20260919.json.

    Script: scripts/audit_phase3a_analogue_fallback.py.