GitHub put Copilot Code Review Balanced first on ReviewBench. I didn’t run the 219 pull requests. The comparison they ask you to use is grounded recall, because augmented recall grows when an agent finds something the golden set missed, and on the 1 Oct board that number is 26%.