Phase 3A retrieval-policy freeze — 2026-09-19

Scope and status

This freezes the candidate-generation policies for the next evaluation pass. It does not claim an official CASMI26 score, hidden-test performance, or Top-5 readiness. The current fold-4 and deterministic reserved readouts are post-freeze diagnostics; neither is a genuinely untouched prospective estimate.

Frozen policies

P0 — conservative baseline

Convert the query precursor m/z to neutral mass using the declared adduct.

Retrieve PubChem structure-index rows within ±10 ppm.

Deduplicate candidates at scorer identity14 level.

Do not use target identity, target-relative mass error, or spectral similarity for candidate generation.

P1 — sodium-wide comparator

[M+Na]+: ±30 ppm.

All other adducts: ±10 ppm.

Same deduplication and leakage rules as P0.

P2 — empty-stage fallback

First retrieve all adducts within ±10 ppm.

Only if the strict candidate pool is empty, append candidates within ±30 ppm.

Preserve retrieval stage as provenance: strict stage 1, fallback stage 2.

Sort deterministically by stage, absolute mass deviation, then identity14.

This is the preferred provisional cost/coverage policy, not a final production promotion.

Required evaluation contract

Every future readout must report:

query count and valid-query count;

spectrum recall and recall at @25, @100, and @1,000;

molecule-any-query recall;

candidate identity count, not raw index-row count;

median, p95, and maximum candidate counts;

formula-gated candidate counts only as a diagnostic when formula is query-provided;

breakdowns by adduct, neutral-mass bin, and formula availability;

retrieval stage and mass deviation provenance;

source/index/configuration hashes and partition definition.

Forbidden signals

precursor_error_ppm and any field derived from the target identity or reference mass are forbidden in candidate generation, fallback decisions, or quality flags. They are target-relative and unavailable at deployment.

Selection rule for the next clean readout

A genuinely pre-registered identity-level partition must be created before inspecting its outcomes. P2 may replace P0 only if it improves query-level and molecule-level recall without unacceptable candidate growth and without an isolated-adduct failure. Until then, retain P0 as the conservative baseline and P2 as the provisional comparator.

Current evidence snapshot

The post-freeze deterministic readout in phase3a-reserved-policy-readout-20260919.json found:

P0: spectrum recall 82.180%, @25 73.955%, molecule-any 93.339%, candidate P50/P95 19/64.

P1: spectrum recall 85.302%, @25 75.802%, molecule-any 93.424%, candidate P50/P95 21/77.

P2: spectrum recall 84.037%, @25 75.348%, molecule-any 93.681%, candidate P50/P95 20/65.

These are stability diagnostics only because the partition was created after earlier cohort inspection.

📑 Contents