Validators Judge. They Don't Help.

Short summary

This article argues that AI code validators should not also fix the issues they flag, because collapsing finding and fixing into one action destroys auditability. The author endorses three emerging practices — verification debt as a concept, inner/outer loop separation for quality gates, and shadow testing for agents — but pushes back on Sonar's 'Solve' stage where the validator writes patches. In regulated environments, a validator that remediates smears judgment and intervention into one event, eliminating the daylight needed for review boards to verify what was caught, who approved it, and what changed.

• •Validators that auto-fix flagged issues destroy audit trails by collapsing judgment and intervention into one event
• •Three endorsed practices: verification debt framing, inner/outer loop quality gates, and shadow testing for agents
• •In regulated shops, separation of finding, fixing, and approval is essential for compliance and accountability

Generated with AI, which can make mistakes.