Pre-registered untouched Phase 3A readout — 2026-09-19

> Audit correction (2026-09-19): “untouched” is not established: 164 registered identities are in previously inspected fold 4. The 462 observations are valid sampled spectra, not distinct molecules; combined MRR is untruncated, not MRR@25. Results remain diagnostic and must not promote a baseline. See review-20260919-1817/AUDIT.md; historical values below are preserved.

Status and scope

A local identity-level partition was registered before this readout:

Source: 5,974 strict-validation identities

Selection: sha256(identity14) first uint32, modulo 7 equals 6

Explicitly excluded: prior post-freeze cohort, modulo 5 equals 0

Selected identities: 739

Fold distribution: 160 / 139 / 152 / 124 / 164

Readout cap: 100 queries per scorer fold

Valid evaluated queries: 462

This is prospective relative to the prior post-freeze diagnostics, but it is still a local catalogue diagnostic—not an official hidden-test or competition estimate. The partition is not certified acquisition-lineage novelty.

Route

The route uses only deployable retrieval fields (adduct, precursor_mz). It excludes the target identity and fold-held-out identities from spectral reference neighbours. Candidate scores are the maximum over up to 25 eligible neighbours of:

cosine(query spectrum, neighbour spectrum) × Morgan-Tanimoto(candidate structure, neighbour structure)

Queries without an eligible spectral neighbour use deterministic neutral-mass ordering as a separate fallback route.

Results

Policy	Valid queries	Spectral rankable	No eligible neighbour	Spectral Recall@25, conditional	Fallback pool recall	Fallback Recall@25	Combined Recall@25	Combined MRR
P0 strict ±10 ppm	462	370	92	91.08%	77.17%	76.09%	88.10%	0.4698
P1 Na ±30 ppm	462	371	91	90.57%	76.92%	75.82%	87.66%	0.4613
P2 empty-pool fallback	462	370	92	91.08%	77.17%	76.09%	88.10%	0.4698

Comparison with prior stability diagnostic

The pre-registered readout is not directly comparable as a leaderboard estimate, but the route-level diagnostic changed materially:

Conditional spectral Recall@25 increased from 86.61% to approximately 91.08% for P0/P2.

Fallback Recall@25 decreased from 89.25% to 76.09%.

Combined Recall@25 increased from 87.13% to 88.10% for P0/P2.

P1 is slightly worse on this partition; sodium widening adds no demonstrated benefit.

The combined result is driven by different route availability and must not be interpreted as evidence of official performance or Top-5 readiness.

Decision

The partition and readout are now registered and complete.

P0 remains the preferred strict baseline for further testing; P1 does not earn promotion from this readout.

The analogue-plus-mass fallback remains a diagnostic baseline, not a promoted competition method.

Learned training remains deferred until maintained regression coverage, full failure-category accounting, and broader leakage/provenance checks are complete.

Artifacts:

Partition manifest: manifests/phase3a-preregistered-untouched-20260919.json

Partition creator: scripts/create_phase3a_preregistered_partition.py

Readout: reports/phase3a-analogue-transfer-preregistered-20260919.json

Readout runner: scripts/diagnose_phase3a_analogue_transfer.py