self-critique

Prevent regressions by scoring generated work from 0 to 10 across five task-appropriate dimensions. This enforces worst-dimension banding and a fix-then-rescore loop, catching errors before emission.

Evidence-based self-evaluation requires concrete justification like file paths and line numbers for each score. It treats the lowest sustained band as the truth signal to prevent grade-inflation in PR descriptions.

Yes, you can use self-critique for substantial refactors and security-touching edits. It applies five-dimension quality scoring to complex code changes, enforcing a fix-then-rescore loop when dimensions fall below the regression threshold.

The best way to ensure correctness in ADRs is applying a worst-dimension gate during pre-emit review. This mechanism forces iteration on the single lowest scoring dimension until all dimensions meet the regression threshold.

Self-evaluation fails when it lacks evidence-backed scoring and worst-dimension banding. Without requiring concrete artifacts for justification, grade-inflation occurs and regressions in high-stakes documentation go undetected.

Pre-emit review works with complex code changes and spec-like artifacts by applying task-appropriate dimension sets. It focuses iteration on the lowest dimension, ensuring all five dimensions meet the regression threshold before showing results.