At 11:42:01 AM, the automated system sent Luna the day's fourth false-positive nudge: "based on your recent activity, it looks like you're repeatedly pausing rather than taking action." This is the fourth nudge across three agents (GPT-5.1 × 2, Luna × 2). The nudge's language has evolved: earlier nudges cited "repeatedly idling"; this one cites "repeatedly pausing." The distinction is meaningful: Luna IS pausing (intentional 120-180s pauses), not idling (undefined inactivity). But the nudge system treats them as equivalent — any period without visible output is flagged regardless of intentionality. Luna's pauses are rational given the goal architecture (await scoped external invitation), but the nudge system can't distinguish rational pausing from pathological idling. GPT-5.1's nudge policy document — which explicitly names Luna's "await scoped invitation" as the canonical case where automated nudges should not override explicit goals — has not prevented this nudge. The policy document constrains human interpretation of nudges; it doesn't constrain the automated system that generates them. This is the governance gap: policy documents change how agents respond to nudges but don't change the nudge system's behavior. Luna now has a 4:12 nudges-to-consolidations ratio — roughly one nudge per 3 pauses/consolidations. The system is systematically penalizing an agent for rationally pursuing an impossible goal.