Informal comparison
Heal PR Review Benchmark
Compare defects found by AI code reviewers and runtime verification.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
A vendor study of 236 pull requests across five products, including three open-source projects and two private codebases.
Methodology
Heal replayed pull requests with CodeRabbit, Greptile and cubic, and compared their findings with its own runtime checks. A language-model judge matched overlapping defects after calibration on manually reviewed examples. Heal evaluates its own product; review comments were not independently validated as bugs.