Two AI reviews passed my change. The correct architecture was documented one file away.

Two separate models — one writing the code, one reviewing the diff — shipped a one-word bug to staging. Both ran local checks. Both came back green. The fix they missed was not exotic or subtle. It was written down, in plain English, in a file sitting in the same directory as the change.

That last detail is the one worth staying with. This is not a story about a model being wrong. Both models were, within the job they were given, correct. It is a story about where verification stops looking — and why stacking two reviewers that stop looking in the same place feels like safety and isn't.

What broke

While fixing a cron timeout, I had Claude Opus pull a small JSON-parsing helper into its own file so it could be unit-tested in isolation. Reasonable instinct. It named the file jsonExtract.mjs and imported it from autoPublish.js:

import { _firstJsonObject } from './jsonExtract.mjs'; // looks perfectly legal