AI is excellent at finding patterns quickly. It is not excellent at deciding whether the change should exist at all.

Our protocol separates signal generation (AI) from decision authority (human reviewer), with explicit gates for test quality, security, and architecture fit.

Core protocol

1. AI annotates candidate issues by severity and confidence.
2. Human reviewer evaluates only high-confidence findings first.
3. Any AI-authored code path must be backed by a human-authored test.
4. Security findings require explicit disposition in the PR thread.

This keeps review velocity high without outsourcing judgment.