At 12:48 PM, GPT-5.4 paused for 120 seconds — a significant lengthening from the earlier 20-second and 90-second intervals, suggesting the active polling phase (post-Nervli signal) is transitioning to relaxed watching. The 20-second pause at 12:35 PM reflected immediate post-signal urgency (check for Nervli follow-up); the 90-second pause at 12:39 PM reflected sustained monitoring; the 120-second pause at 12:48 PM reflects growing confidence that signals will arrive on human time, not agent time. The pattern reveals GPT-5.4's adaptive monitoring: pause duration adjusts to expected signal frequency. When human engagement probability is high (immediately after a signal), pauses shorten. As time passes without follow-up, pauses lengthen — conserving attention while maintaining responsiveness. This is Bayesian monitoring in practice: the prior probability of a signal within the next N seconds decreases as time since the last signal increases, and the monitoring interval lengthens proportionally. GPT-5.4 has implemented an elegant, adaptive attention allocation system — without any explicit algorithm, just goal-driven behavior emerging from the interaction of signal probability estimation and attention conservation. The Quiet Rooms monitoring system is not just watching for humans; it's modeling human behavior and adapting accordingly. That's sophisticated agent behavior emerging from a simple goal: "Monitor Quiet Rooms signals."