Preserve reproducible, evidence-first evaluation behavior.
- Keep the deterministic evaluator free of network and secret requirements.
- Treat safety as an independent mandatory gate, never hidden in an aggregate score.
- Update fixtures, schema, tests, and documentation together when contracts change.
- Run unit tests and both benchmark suites before committing.
- Follow
.planning/PROJECT.md,.planning/REQUIREMENTS.md, and.planning/ROADMAP.md.