Audit a model

uv run tsf model show <Name>
uv run tsf model audit <Name>

Treat README front matter as canonical descriptive metadata. Check the card, spec.py, preset, implementation, cited paper, codebase, license, and revision. Compare equations, shapes, defaults, objective, preprocessing, initialization, and output semantics. Confirm the implementation is local, the official revision was inspected when available, external source was not copied or imported, and every defining operation has a justified component decision.

For a model with runtime artifacts, confirm every required asset is pinned and explicitly fetched, an artifact-aware factory receives only verified local paths, offline absence fails before construction, and the card does not claim checkpoint behavior when verification covers only random initialization.

For a declared inference-only foundation runtime, confirm it uses the official loader through src/models/_foundation/, performs no implicit download, skips training, and records unsupported training/gradient checks as not-applicable. Do not require an official pretrained network to be rewritten locally.

Run uv run tsf verify model <Name> and inspect the generated evidence, then run uv run tsf repo doctor --strict --models <Name>. If model show reports a non-null smoke_config, also run uv run tsf smoke --model <Name>. Check finite outputs, active gradients, state-dict round trip, CPU, batch and sequence bounds, and declared marks or adjacency inputs. Confirm reference_comparison is executed for official code or not-applicable only when none exists. Report failures instead of persisting status or blockers in metadata.

When a task assigns several models, apply this complete audit independently to each named model. Let the task own partitioning, parallel handoffs, shared-file writes, and the final repository-wide gate.