Skip to main content

4.6 Multi-instance and multi-pass review architectures

Self-review is structurally weak​

A model that generated the code retains its reasoning context and is therefore less likely to question its own decisions in the same session. An independent review instance with no prior reasoning context is more effective at catching subtle issues than either a self-review instruction or extended thinking.

generate ──► [instance A] ──► code
review ──► [instance B] ──► findings (B never saw A's reasoning)

Multi-pass review​

For a 14-file PR producing inconsistent depth and contradictory findings:

per-file pass : file_i alone -> consistent local depth
integration pass: all files, data-flow -> cross-file issues

The root cause is attention dilution, so the fix is decomposition.

Distractors on the multi-pass item
  • A larger context window does not fix attention quality.
  • Requiring consensus across three full-PR runs suppresses real bugs that are only caught intermittently.
  • Making developers split PRs shifts the burden without improving the system.

Confidence-annotated verification passes​

Run verification passes where the model self-reports confidence alongside each finding, then route review attention accordingly. Note the boundary: self-reported confidence is fine for routing review capacity; it is not a substitute for a decision gate (see Domain 5.2).

Quiz​

  1. Question 1

    A pull request modifies 14 files. Your single-pass review produces detailed feedback for some files and superficial comments for others, misses obvious bugs, and contradicts itself — flagging a pattern in one file while approving identical code elsewhere in the same PR. How should you restructure the review?

  2. Question 2

    A CI step has Claude generate a database migration and then, in the same session, review its own work for correctness. It consistently approves scripts that a human reviewer later finds contain an unguarded backfill over a live table. Adding "review critically and assume there is a bug" to the review turn changed nothing measurable. What is the structural fix?

  3. Question 3

    You already split large reviews into per-file passes plus an integration pass. A reviewer points out that the integration pass now re-reports issues the per-file passes already raised, and that the two passes sometimes disagree about severity on the same line. What should change?

  4. Question 4

    You want to route review findings: post the reliable ones automatically and send the rest to a senior engineer. The proposal is to have the model report a confidence value with each finding and route on that. What makes this workable?