| name | audit-code-artifact |
| description | Audit scientific implementation, exported-model equivalence, dependency completeness, manifests, packaging, and isolated inference. Use when claims depend on code, trained artifacts, evaluators, or deployable entry points. |
Audit Code and Artifact
Inspect existing files and test outputs read-only. Do not execute the scientific
pipeline or manufacture missing validation evidence.
Checks
- Trace critical values across data loading, preprocessing, fitting, export, and
inference. Check defaults, branches, feature order, transforms, and manifest
declarations against the scientific protocol.
- Verify that decision-relevant tunable parameters live in machine-readable
configuration separate from the main implementation logic. Map each
decision-relevant parameter from its stable name, value, unit, and provenance
to the code that consumes it. Flag duplicated or unexplained values when they
prevent a reviewer from determining which setting actually ran.
- Require existing numeric evidence that the reference pipeline, exported model
or raw weights, and final entry point produce equivalent predictions on fixed
samples within a stated tolerance.
- Require an existing clean-directory or evaluator-like smoke test using only
collected artifacts and declared dependencies.
- Check for undeclared local modules, absolute workspace paths, environment
variables, auxiliary files, incompatible versions, and entry-point assumptions.
- Separate implementation correctness from scientific adequacy: a correctly
packaged artifact does not establish that its model or validation is suitable.
For a bounded parallel review, use code-reviewer for concrete implementation
defects and repo-scout only when imports or artifact dependencies must first be
mapped. Ask for evidence and candidate findings, not a verdict.
