AI leaders warn we could build a system that, as Sam Harris put it, "could, at a certain point, no longer care what we want" — autonomous agents that self-improve and slip human control. That's the control problem, and it's the whole reason the UN Security Council is now hearing about AI.
You can't trust a model to police itself. GARRISON puts control OUTSIDE the model — a boundary it can't see, bypass, or falsify — with a kill switch and tamper-proof alerts. It's the control problem applied to the AI we actually run in production today. It does not claim to contain hypothetical superintelligence — and being honest about that is the point.