A benchmark is a controlled argument about capability. Work is less obliging. Inputs arrive incomplete, instructions conflict, and the cost of a plausible error may matter more than the average score.
The evaluation that follows deployment therefore deserves as much care as the evaluation that preceded it. Measure corrections, abandoned outputs, time recovered, and the cases quietly routed back to a person.
Headline numbers can identify a promising system. Only ordinary use can establish whether it deserves to stay.