Lesson 3 of 4 in Context Management & Reliability
5.3 · Multi-instance & multi-pass review
Self-review is weak because the same reasoning that generated the output rationalizes it during review. Independent instances catch what self-review misses. This lesson covers when to use independent review, multi-pass patterns for large artifacts, and confidence-tagged findings for routing.
Self-review is weak
A model instance that just generated content retains its reasoning context. Asking the same instance to critique its own work shares those assumptions. It's less likely to question its own decisions than a fresh instance would.
Independent review
A second Claude instance without prior reasoning context is more effective at catching subtle issues. This is the standard for high-stakes outputs: security review, contract review, production code review. The reviewer starts fresh.
Multi-pass for large artifacts
For big multi-file or multi-section artifacts, one pass over the whole thing dilutes attention. Split into per-artifact local passes plus a separate cross-artifact integration pass. Local passes catch details; the integration pass catches consistency.
50-page contract → per-section reviews for each clause → cross-section integration pass for consistency and conflicts
Verification with self-reported confidence
In review passes, have the model report confidence alongside each finding. High-confidence findings can auto-apply; low-confidence findings route to human review. Confidence calibration turns a review into a routing decision.
Takeaways
- Self-review shares context with the generator — weak
- Independent instances catch more
- Split big reviews into per-artifact + integration passes
- Confidence-tagged findings enable calibrated routing
Exam traps
Practice scenario
A team asks the SAME Claude instance that just generated a 500-line service module to review its own code for bugs. The review is superficial and misses real issues. What's the correct architecture?