Put a cheap model in charge of the small calls — stop now, verify first, change approach — and the usual evidence that it works is that its judgements are accurate. Here one terminal-agent loop was held fixed, only the controller varied, and the arm nobody usually runs was added: a placebo that interrupts at exactly the same moments with choices drawn from the real controller's own answer frequencies. The controller's judgement really was better. It cost 19% more per attempt for the same 73.0% pass rate, and ran 22% dearer than the placebo that matched it for nothing.