Reproducibility & Datasets

A result you can't reproduce isn't evidence. The factory is built so its inputs are pinned and its public datasets are versioned — so a verified badge means the same thing tomorrow as today.

Fork it and run it yourself — the output is recorded, not asserted (e.g. HG002 Ts/Tv ≈ 2.0).

Reproducible by construction

• Pinned build — the docs site and the platform pin their generators; CI runs the real gate (lint · platform tests · registry validation) with no || true, and the strict docs build fails before publishing anything stale.
• Deterministic lineage — governed runs carry a manifest (inputs → capability chain → serving → composed honesty tier → a deterministic lineage hash), so a result traces to exactly what made it. See Governance & Lineage.

Public-datasets manifest (version-pinned)

Every public dataset used across the engines and agents is pinned to a release, so cross-omics joins stay reproducible over time. Data itself is never committed — only its provenance.

Dataset	Version / release	Used by
GRCh38 primary assembly	Ensembl release-112	foundation / genomics
HGNC complete set · HPO · Cell Ontology · Reactome	current releases	BioKey resolver, phenotype/pathway
ClinVar (GRCh38)	weekly VCF	E1 annotation · ACMG secondary-findings (verified)
AlphaMissense hg38	v1 (2023)	missense pathogenicity
ChEMBL	chembl_37 (SQLite)	therapeutic discovery reference
CPIC (diplotype/recommendation/allele tables)	API snapshot	pharmacogenomics
Orphanet (Orphadata product 1/4/6)	current en_	rare disease
CIViC (evidence/variant/gene)	nightly	precision oncology
10x pbmc3k	scanpy example	single-cell compute (verified)

The full manifest and a resumable downloader live in the neutral repo's central public-datasets store; only code and docs publish — data, weights, and secrets stay local.

Note

Version pins are the difference between "it worked once" and "it reproduces." Where a capability is marked verified, the exact dataset + version it was proven against is recorded on Citations & Evidence.