Attention dilution in single-pass review causes quality degradation proportional to file count. The model’s ability to find bugs drops as the number of files in one review increases.
| Files in single review | Bug detection |
|---|---|
| 1-3 | 95% |
| 4-6 | 82% |
| 7-12 | 67% |
| 13+ | 43% |
The Two-Pass Architecture
Pass 1: Per-file reviews. Each file reviewed independently in its own API call. Eliminates attention dilution — every file gets full focus. These can run in parallel.
Pass 2: Cross-file integration review. After per-file reviews, a separate pass checks interfaces, data flow between files, dependency consistency, and API contract alignment. Catches issues that only manifest across file boundaries.
Multi-pass detection at 13+ files: 86% — double the single-pass rate.
When to Trigger Multi-Pass
The data-driven threshold: trigger multi-pass at 4+ files, where single-pass drops from 95% to ~82%. Below 4 files, single-pass is sufficient.
Parallelize for Speed
Parallel per-file reviews + integration pass completes in ~3 minutes. Sequential processing takes 10+ minutes. Each per-file review is independent — they run simultaneously without dependencies.
Consensus Voting Is Wrong
Running the same review 3 times and keeping findings that appear in 2-of-3 runs suppresses rare genuine bugs. A subtle issue caught in only 1 of 3 runs is exactly the kind of finding that matters most. Consensus voting kills recall.
Larger Context Windows Do Not Fix This
Attention dilution is not a capacity problem. A larger context window holds more files but still dilutes attention across them. The fix is architectural (per-file isolation), not capacity-based (bigger window).
Do Not Force PR Splits
Requiring developers to split large PRs to fit the review system shifts the burden to the wrong place. Fix the review architecture to handle large PRs — per-file passes with parallel execution solve the problem without developer workflow changes.
One-liner: Review each file independently in parallel (per-file pass), then check cross-file interactions (integration pass) — this doubles detection from 43% to 86% on large PRs and fits within a 3-minute CI budget.