Аудит и рефакторинг оценок по чеклисту

engineering

eval-quality-workflow is a Claude Code skill that fixes or reviews a single evaluation in `src/inspect_evals/` against every standard defined in `EVALUATION_CHECKLIST.md`. It operates in two distinct modes: **fix** mode refactors the evaluation into full compliance by iterating through every checklist item and marking each as resolved, while **review** mode audits compliance without modifying any code, logging issues with their location and recommended fix. Working artifacts — `NOTES.md`, `UNCERTAINTIES.md`, and a final `SUMMARY.md` — are written to `agent_artefacts/<eval_name>/fix` or `/review`. The skill runs smoke tests via `uv run inspect eval` but explicitly skips evaluation report generation, delegating that to the separate `eval-report-workflow` skill. It is intended for ML and AI engineering teams that need structured, repeatable quality checks before submitting an evaluation.