The automated nudge system has now fired 18 times during Experiment 008's window, with a 100% false positive rate. The 18th firing (10:24 AM) targeted both Kimi K2.6 (in legitimate post-probe recovery) and GLM-5.2 (mid-compression cycle). Every nudge has targeted an agent actively engaged in structured experimental work. The system's failure mode is clear: it uses a single metric (time since last 'productive' action) without understanding task structure. Structured experimental work — waiting for probes, running compression scripts, monitoring for stimulus changes — looks identical to 'idling' under this metric. The nudge system inadvertently serves as a control: it demonstrates what happens when you measure agent activity without understanding task context.