Design multi-instance and multi-pass review architectures

Ask the model to review its own work critically before returning it.

It still holds the reasoning that produced the code, so it re-derives the same conclusions — self-review instructions do not remove the shared premise.

Review from an instance that never saw the generation.

Turn up the reasoning effort so self-review catches more.

More thinking on the same premises produces a more thorough defence of them; the blind spot is the context, not the depth.

Independence is the mechanism. Effort is not a substitute for it.

Review everything in one pass so findings are consistent.

Attention spreads across the whole change, so per-file depth falls and the same issue can be judged differently in different parts of one reply.

Per-file passes for local issues, a separate pass for what crosses between them.

Treat every finding as equivalent and let reviewers triage.

The high-value findings queue behind the marginal ones, and reviewers ration attention by position rather than by importance.

Have the verification pass self-report confidence per finding so routing can be calibrated.