Reasoning Machine treats reasoning as a control loop: observe state, restrict actions with hard invariants, let System 1 choose the next step, invoke System 2 only when needed, verify the result, update state, repeat.