AI review benchmarks score the bot. Nobody scores the reader

This week GitHub published ReviewBench, an open benchmark for AI code reviewers. It is careful work: 219 pull requests from 187 open-source repositories across 19 languages, picked so their size match