# Biology benchmark evidence supplement

This supplement supports an exploratory article about coverage and metric definitions. The four pilot investigations remain pending human review. These results do not establish statistical superiority, biological mechanisms, novelty or performance across complete benchmark suites.

Start with the [quantitative audit](evidence-audit.md), [aggregate results](reproduction/results.json) and [reproduction instructions](reproduction/README.md). [Evaluator provenance](reproduction/evaluator-provenance.json) records exact hashes and pinned source revisions. The MFASS ordering reversal concerns a research diagnostic recall calculation; the database's precision-at-100 ordering does not reverse.

## Reproduction access

The included standard-library command checks arithmetic using the supplied aggregate values:

```sh
python3 reproduction/reproduce.py aggregate
```

This checks recall fractions, NDCG arithmetic and the relative MSE reduction. It does not authenticate the source observations, reproduce inference or independently verify top-100 membership.

Full numerical recomputation requires acquiring the exact original source files and unpublished derived tables/prepared artifacts referenced by the campaign records. Some artifacts have no public retrieval URL. This package alone is therefore insufficient for a fresh full recomputation. Keep a private artifact-ID-to-filename resolver outside the supplement and follow the [full reproduction instructions](reproduction/README.md). No raw assay data, biological sequences, row-level observations, private resolver, credentials or model-session logs are distributed here.

## Preserved investigation records

The following four exports are exact byte-preserving copies, including their original frozen plans, execution receipts, aggregate usage and review status. Plan prose is not evidence that an operation ran; use the executed receipts. In particular, the mRNABench expected-observation prose reverses the comparison labels; the quantitative audit explains the actual receipt direction.

- [MFASS v2](mfass-v2-discrepancies-campaign-fd29c4413eaf4e24.json)
- [FLIP2 Rhomax](flip2-discrepancies-campaign-fd29c4413eaf4e24.json)
- [mRNABench designed Sample MRL](mrnabench-discrepancies-campaign-fd29c4413eaf4e24.json)
- [ProteinGym AMFR](proteingym-amfr-discrepancies-campaign-fd29c4413eaf4e24.json)

[Packaging checks](packaging-check.json) record file hashes, byte-copy verification and the scope of the content review. The copy was scanned recursively for personal filesystem paths, credential patterns, biological sequence strings, raw observation arrays and session-log fields. The campaign `rows` fields are aggregate counts. Their `execution.reasoning` fields contain session metadata and token counts, not transcripts or private chain-of-thought. Public source links retain their original provenance; no public location is asserted for unpublished code or data.

## Additional article-preparation check

After the frozen pilot, a descriptive [feature-coverage check](reproduction/feature-coverage.json) found one recorded prepared-input length across all 15,003 mRNA test rows and two recorded GC bands. Length stratification cannot provide a between-group comparison in this representation. No GC-specific outcomes were tested. The [check script](reproduction/check-feature-coverage.py) requires the same private resolver and authenticates the table hash before counting feature metadata. This check is separate from the pilot's 22 registered attempts and does not alter its frozen plans or reports.

The [baseline metadata](reproduction/model-baselines.json) records the exact model descriptions and feature configurations from the original hash-verified reports. FLIP uses amino-acid composition with fixed ridge regression; the mRNA control uses sequence composition with training-only RidgeCV. Neither composition control uses pretraining. These metadata clarify the original runs and do not represent new fits.