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

📑 Contents

Prospective Phase 3A analogue-transfer accounting — 2026-09-19ScopeP0 strict ±10 ppmP0 by adduct — combined Recall@25P0 by neutral-mass bin — combined Recall@25P1/P2 comparisonPolicy cleanupProvenance

Prospective Phase 3A analogue-transfer accounting — 2026-09-19


> Audit qualification (2026-09-19): the original narrative below is superseded where it calls the partition untouched, MRR scorer-compatible, or route/category counters literal target-presence/cap counts. 164 registered identities overlap the earlier fold-4 diagnostic; MRR is untruncated; the 92 no-positive-reference cases include absent targets. These are sampled spectrum-level diagnostics. See review-20260919-1817/AUDIT.md before using this report for decisions. Raw JSON is preserved as historical output, not automatically authoritative semantics.


Scope


This is a rerun on the already registered partition:


manifests/phase3a-preregistered-untouched-20260919.json


It uses only adduct and precursor_mz for retrieval, excludes target and fold-held-out identities from spectral references, and is local prospective evidence only. It is not an official score, hidden-test estimate, acquisition-lineage novelty certification, or Top-5 evidence.


P0 strict ±10 ppm


  • Valid queries: 462
  • Target absent from structure pool: 39
  • Target present but no eligible positive spectral reference route: 92
  • Spectral route produced a ranked candidate list containing the target: 352
  • Fallback route ranked the target: 71
  • Target outside the 1,000-candidate reporting cap: 0
  • Candidate pool size P50/P95: 16 / 63
  • Combined Recall@25: 88.10%
  • Combined MRR: 0.4698

    • The category counters are route observability counters and are not mutually exclusive in every implementation detail: the structure-pool miss count can overlap with an unavailable-reference route, and eligible_reference_no_positive_candidate_score means no candidate received a non-empty fingerprint score map—not that all candidate scores were mathematically zero. The raw JSON is authoritative.


    P0 by adduct — combined Recall@25


  • [M+H]+: 89.03% (383 queries)
  • [M+K]+: 100.00% (7)
  • [M+NH4]+: 60.00% (10)
  • [M+Na]+: 72.41% (29)
  • [M-H]-: 95.65% (23)
  • [M-2H2O+H]+: 100.00% (4)
  • [M-H2O+H]+: 100.00% (5)
  • [M+Cl]-: 100.00% (1)

    • P0 by neutral-mass bin — combined Recall@25


  • 0–250 Da: 85.20% (223 queries)
  • 250–500 Da: 91.98% (187)
  • 500–750 Da: 82.05% (39)
  • 750+ Da: 100.00% (13)

    • P1/P2 comparison


  • P1 Na ±30 ppm: combined Recall@25 87.66%, combined MRR 0.4613; candidate P50/P95 19 / 63.
  • P2 explicit empty-pool fallback: combined Recall@25 88.10%, combined MRR 0.4698; candidate P50/P95 16 / 63.

    • P0 and P2 are identical on this partition. P1 widening is not promoted.


    Policy cleanup


    P2 is now represented with separate candidate-pool and reference-window fallback configuration fields. Its observed equality with P0 does not establish universal equivalence; future changes must report the two stages independently.


    Provenance


  • Structure index rows: 231,749
  • Partition manifest: manifests/phase3a-preregistered-untouched-20260919.json
  • Accounting JSON: reports/phase3a-analogue-transfer-preregistered-accounted-20260919.json
  • Runner: scripts/diagnose_phase3a_analogue_transfer.py