Lesson 4.6Lesson 24 of 30
Multi-instance and multi-pass review
Design multi-instance and multi-pass review architectures
Key points
- 1Self-review is weak: in the same session the model keeps the reasoning that produced the code, so it tends to confirm its own decisions instead of questioning them.
- 2A second, independent instance that gets only the artifacts (diff, spec, test results), not the generator's conversation, catches subtler issues than "review your work carefully" or extended thinking in the same session.
- 3Split a large multi-file review into per-file local passes (each file gets full attention) plus a cross-file integration pass (interfaces, data flow, contracts). One giant pass dilutes attention and produces contradictory findings.
- 4Add a verification pass in which each finding carries evidence and a self-reported confidence. Use it to route: confident findings go straight to the PR, uncertain ones to a human queue. Calibrate thresholds against labelled outcomes.
- 5Running the same reviewer prompt N times mostly repeats the same blind spots, and keeping only findings most runs agree on drops the real bug only one run caught.
What the exam guide checks here
Objective wording from the exam guide (English).
- 4.6.K1Self-review is weak: the model keeps its generation reasoning
- 4.6.K2Independent instances catch more than self-review instructions or extended thinking
- 4.6.K3Multi-pass: per-file local passes plus cross-file integration passes
- 4.6.S1Use a second independent instance without the generator's context
- 4.6.S2Split multi-file reviews into per-file and integration passes
- 4.6.S3Verification passes with self-reported confidence per finding for routing
Lock it in
Answer scenario questions on this topic. Every option comes with an explanation.
Practise 4.6 · 16 questions