Model reference functions

Model reference reads use the catalog revision effective at the requested UTC time; omitted time means now. Publications are complete, immutable snapshots, so a later price change does not alter the revision selected for an earlier session. The bundled catalog seeds an empty database once; subsequent updates use these functions without a code release.

Function id	Claim	Handler	Result
models.lookup.run	None	yoke_core.domain.handlers.model_reference	One model, researched, revision_id, effective_at
models.get.run	None	same handler	Catalog or model plus revision provenance; accepts at or revision_id
models.validate.run	None	same handler	Validated record
models.diff.run	None	yoke_core.domain.handlers.model_reference_publication	Complete-catalog diff against the latest scheduled revision
models.revisions.run	None	same handler	Ordered revision history
models.publish.run	None	same handler	New revision and diff; org admin, expected base, source note, optional future effective_at
models.restore.run	None	same handler	New revision copied from an earlier revision; same publication guards
models.level_proposal.run	None	yoke_core.domain.handlers.model_level_proposal	Proposed universe levels: generated or authored changes, the resulting levels, unverified options, unplaced_models; refuses an option its model's published effort or context window contradicts

Records hold facts about models only: identity, successor (replacement_model_id), published reasoning_efforts and context_window_tokens, prices, subscription rules, sources. A candidate carrying the retired proposed_tier, tier_evidence, or tier_provisional keys is refused with tier_classification_retired. Which model a worker launches is decided by the execution levels; models.level_proposal.run proposes level changes and writes nothing, and universe.levels.set stores the operator-approved document, refusing the same published-value conflicts.

The $models skill teaches source review and CLI recipes. yoke models <command> --help gives each payload. Model availability and reasoning levels still come from harness-native discovery.