GEN-PREP — shared generation contract for arXiv-inspired challengers

Status: design-only, untrained, reference-only. Covers MS-BART/MARLIN/DiffMS/MADGEN-inspired branches without importing their code, weights, tokenizers or datasets.

Shared deployment boundary

Every generator receives only deployable inputs available at inference:

spectrum-derived features or predicted fingerprint logits;

permitted precursor m/z, adduct/polarity and metadata missingness;

predicted formula hypotheses only when the formula route is explicitly selected;

retrieved/predicted scaffold only for scaffold-completion routes.

Ground-truth structure, hidden-test formula, target identity, target-derived precursor error and held-out spectral prototypes are oracle/audit-only and must never enter deployable generation.

Route contracts

MS-BART-inspired fingerprint-to-structure

Structure pretraining, if justified, uses eligible training structure pairs only.

Adaptation uses OOF predicted/noisy fingerprints, never in-sample perfect fingerprints as the sole result.

Report oracle-fingerprint and deployable predicted-fingerprint controls separately.

MARLIN-inspired formula-free generation

Inputs: predicted fingerprint representation plus permitted precursor mass/adduct metadata.

No hidden true formula.

Report exact new scorer identities, in-pool recovery, valid-SMILES rate and final MRR after shared reranking; validity alone is insufficient.

DiffMS-inspired formula-conditioned generation

Inputs: a declared ranked set of predicted formula hypotheses.

Ground-truth formula is an oracle diagnostic only.

Report formula retention/coverage, formula-miss strata, invalid generation and exact identity recovery.

MADGEN-inspired scaffold completion

Inputs: retrieved or independently predicted scaffold, never held-out target scaffold.

Separate scaffold retrieval failure, invalid completion, wrong identity and ranking failure.

No fixed scaffold quotas or route slots.

Common output contract

Each query emits up to 25 ranked candidates after scorer-identity deduplication. For every candidate retain:

generated SMILES and validity status;

scorer identity when valid;

route and input provenance;

rank, score and upstream rank if applicable;

duplicate identity flag;

failure status and runtime.

Required aggregate metrics:

full-cohort MRR@25, Recall@25, Top-1, Top-5;

valid-SMILES and unique-identity rates;

candidate/pool coverage and covered-pool conditional MRR;

newly correct identities not present in the upstream pool;

formula/scaffold/scorer-identity failure strata;

runtime, memory and timeout rate.

Shared exposure and evaluation policy

Use the scorer-compatible molecule-grouped split. Keep structure pretraining, target spectra, spectral prototypes and learned features out of held-out B/C routes. A prior exposed panel is diagnostic, not an untouched claim. Compare every route on the same query cohort, candidate pool and ranking budget; preserve all unavailable and zero-candidate queries in the denominator.

Acceptance gates

1. Static input/output and scorer-identity fixtures pass.

2. Oracle controls are clearly labelled and never mixed with deployable scores.

3. Predicted/noisy inputs are generated OOF or from an independently frozen upstream model.

4. Beam/sample outputs are deduplicated by scorer identity before ranking.

5. Route failure, pruning loss and ranking loss are separable.

6. Promotion requires marginal full-cohort MRR@25 gain over FP/retrieval controls, not merely valid structures or novel-looking outputs.

Deferral rule

Do not select an architecture or train a generation branch until FP-01/02 and failure strata show generation can add candidates or improve ranking within the approved compute budget. No external assets are approved by this contract.

📑 Contents