Behavior-based code review evals

The c-CRAB benchmark evaluates code-review agents by whether review comments lead to correct fixes that pass tests, rather than by textual similarity to human reviews. Why it matters: For agent-generated pull requests, plausible review prose is not enough. Review systems need executable checks or testable claims that verify whether a suggested change actually improves the code.

Try this: When using an AI reviewer, ask for reproduction steps, failing tests, or a concrete patch-verification path instead of accepting narrative comments alone.