Direct answer
Loop review is a controlled review cycle for AI-generated outputs: compare the output to the locked source, identify drift or gaps, patch only the affected parts, then review again. It improves quality because it treats review as a loop, not a single pass.
How to use it
- Review against source truth first, not taste.
- Mark gaps, overclaims, unsupported additions and tone drift.
- Patch the smallest necessary section.
- Re-run the output through the same checklist.
- Stop when the acceptance criteria are met.
Example
If an article loses the caveat that a framework is diagnostic, not predictive, the patch is not a full rewrite. Restore that caveat, check adjacent sections, then lock the corrected version.
Caveats and limits
Loop review fails when the reviewer keeps changing the brief mid-loop. Separate new ideas from defects.