Designing eval harnesses that actually catch regressions

A short essay placeholder. Real content goes here before launch.