Review recall
Review recall is the share of real defects in a change that a review catches: caught defects divided by all defects present.
Why it matters
Recall is the number that matters for safety, because every defect the review misses can become an escaped defect. It is hard to measure on real pull requests, since nobody knows the full list of bugs a change contains. Benchmarks solve that by planting known bugs and counting which ones a reviewer names.
DiffDojo's living benchmark does this for AI models on the same pull requests people practise on. As of September 2026, five models reviewed 202 PRs, and overall recall ran from 86% to 98%, from 73% to 100% on security bugs. Human reviewers on DiffDojo are graded the same way: a review passes when it catches the planted bug with a comment on the right lines that says what breaks.
Recall also depends on the kind of bug. On code that does not do what the ticket asked, no model in that run went above 96%, because there is no wrong line to point at, only a requirement to compare the code against.
Sources
See also
- Review precision
- Escaped defect
- Worked example: IDOR: any logged-in user can read every other user's order in Node, an ownership check missing from one route