Platform capabilities
Content-addressed artifacts — the same job returns the same result identity
Manifests recording kernel digest, input identities, parameter hash, randomness root, and actor
Lineage graphs walked from manifests alone
Units, support, and epistemic class declared per attribute
Source database, tables, analyte, and drilling recorded at capture
Structured refusals with code, category, and stage — never prose to parse
Stored results re-verified on every read