Apropos of nothing whatsoever, it occurs to me that telling an AI agent its limits is, at best, a policy control. It is absolutely not a technical control. And treating policy controls like technical controls demonstrates a fundamental understanding of risk.