Multiple lifecycles intersect
A model-breeding program manages more than models. Source research, datasets, contracts, evaluation suites, prompts, policies, model packages, and public documentation each evolve at different rates. Changes must remain traceable across those layers.
Lifecycle objects
| Object | Version trigger | Retirement trigger |
|---|---|---|
| Source report | New original or corrected file | Never overwrite; supersede with note |
| Curated guide | Editorial or factual change | Redirect or archive old route |
| Dataset manifest | Added, removed, relabeled, or relicensed records | Expiry, consent withdrawal, revocation |
| Contract | Schema or semantic change | Clients migrated and compatibility window closed |
| Evaluation suite | New risk, drift, or benchmark revision | Retain for historical reproducibility |
| Model package | Any artifact or metadata change | Replaced, revoked, unsupported, or uneconomic |
| Policy | Governance decision | Explicit supersession only |
| Search index | Content build | Rebuild from authoritative sources |
Source preservation
Preserve original reports and hashes. Curated content can correct or qualify claims without rewriting the source. Link every guide to its source reports and record the editorial build date.
Evaluation versioning
A score is meaningful only with a suite version, data manifest, runtime, and policy. When the suite changes, do not overwrite old scores. Re-evaluate active packages and mark old evidence stale.
Revocation propagation
FUNCTION descendants_affected_by(revoked_object)
direct <- LINEAGE_INDEX.children_of(revoked_object)
affected <- TRANSITIVE_CLOSURE(direct)
RETURN affected + DEPENDENT_CONTRACTS_AND_RELEASES(affected)
END FUNCTIONA revoked dataset, signer, dependency, or policy can affect many descendants. Precompute lineage indexes so incident response does not depend on ad hoc searches.
Documentation as control
Operational guides, runbooks, and model cards are part of the safety system. Version them with code and packages. A release that changes behavior but leaves documentation stale creates hidden operational risk.
Retention
Define separate retention for raw user data, telemetry, evaluation evidence, model artifacts, and public research. Preserve the minimum necessary for reproducibility and accountability while respecting deletion obligations.
Source reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.