GLACIER-PREP — bounded forward reranker contract

Status: design-only, paper-inspired, untrained. This is not a reproduction of arXiv:2606.29161 and uses no paper code, checkpoint or external asset.

Purpose

Test whether a forward spectrum-consistency score improves ordering inside a high-recall candidate shortlist. This branch is allowed only after FP-01/02 shows candidate recall is adequate and ranking is the limiting stage.

Inputs

For each query spectrum and candidate structure:

query MS/MS peaks after the frozen cleaning policy;

permitted precursor m/z, adduct, polarity, instrument, collision-energy value and missingness;

candidate scorer identity and structure representation derived from the competition-only pool;

candidate rank/score from the upstream route;

no hidden target formula, target-derived precursor error or query identity feature.

Interface

forward_score(query, candidate) -> {score, status, diagnostics}

The implementation must be independently written and return a bounded score or an explicit failure status. A failed candidate preserves its upstream score; it must not silently disappear from the denominator.

Required status strata:

scored

invalid_candidate

unsupported_adduct

unsupported_charge_or_formula

empty_or_unusable_query

timeout

runtime_error

Feature contract

The first bounded implementation may use only declared, reproducible features:

predicted/structure-derived fragment masses only where the chemistry policy permits them;

query-to-predicted fragment matching with fixed tolerance and deterministic tie-breaking;

matched and unmatched intensity summaries;

precursor/adduct consistency;

collision-energy compatibility or explicit missingness;

upstream score and rank only as separately reported fusion features.

Do not use target identity, hidden test truth, source-library priors, or post-hoc failure indicators as features. Any learned forward predictor requires OOF upstream features and a pinned train/development/report partition.

Evaluation gates

1. Verify candidate retention before reranking: pool coverage and candidate Recall@100/1000.

2. Compare upstream-only, forward-only and fixed fusion on the exact same shortlist.

3. Report full-cohort MRR@25, Recall@25, Top-1/5, covered-pool conditional MRR, invalid/failure strata, shortlist retention, runtime and peak memory.

4. Preserve per-molecule candidate scores/ranks and query/candidate provenance.

5. Promote only with paired molecule-level improvement and no severe intended-regime regression.

Synthetic checks

Before any model or chemistry implementation, validate deterministic tie-breaking, failure preservation, no target-derived feature access, and denominator accounting with synthetic fixtures. See scripts/validate_arxiv_prep.py.

Deferral rule

Do not train or run this branch merely because the paper exists. If FP-01/02 has poor candidate recall, improve retrieval instead. If upstream ranking is already strong, require a preregistered marginal-gain hypothesis before spending GPU time.

📑 Contents