Design high-quality benchmarks that go beyond recall to evaluate how frontier models truly reason, act, and judge.