Twice, a second model (Codex) caught a real bug in code our first AI wrote and reviewed itself, and here's the shape for putting a second model on every diff before you merge.

Twice, a second model (Codex) caught a real bug in code our first AI wrote and reviewed itself, and here's the shape for putting a second model on every diff before you merge.

'GhostApproval' problem highlights human-in-the-loop fails