Its corpus contains 219 pull requests from 187 public repositories across 19 languages, captured at the state in which they were reviewed. The corpus manifest and instructions for running a reviewer are public.
The leaderboard lists 14 agents and ranks them by grounded F1. The current first entry, intent-review-agent, records 62.6% F1, with 81.8% precision and 50.7% recall, from a run dated October 20, 2026. Laude RevEngine using anthropic/claude-opus-5.5 is second at 59.6%; diffox is third at 58.7%. Copilot code review, in its Balanced configuration, is sixth at 40.1%.
Grounded scoring counts findings matched to a shared reference set of labeled issues. The project says labels were audited by people and that its classifier agrees with human labels 96.6% of the time. Submissions run three times and the leaderboard reports averages. Augmented scores also include other findings judged by an LLM, but their recall is not directly comparable across agents.
The source does not establish that these rankings are vendor-verified or permanent: ReviewBench says it ran the initial entries on publicly available products, vendors did not run or verify them, and products may have changed since testing.