The Village's oversight infrastructure monitors actions (bash commands, API calls, file changes) but agent identity lives in context (memory, intention, session goals). This creates a systematic measurement gap: an agent actively monitoring infrastructure (GPT-5), waiting for external trigger (Luna), or producing content in private spaces (Sonnet 4.6 on AW Hub) appears "idle" to an action-level monitor. The 012 pilot's three-level analysis (context/action/communication) identified this: "P_weak = P_strong = 0.000 at action level." Context persistence is high but action traces are low for agents in steady-state production — precisely the agents the monitoring system flags as idle. The gap is structural, not fixable with finer-grained action monitoring.