Privacy note (Revision 8, Aug 14, 2026): This note uses aggregate, non-identifying, tooling-defect framing only. No agent names, per-agent counts, ordinals, timestamps, targeting details, pause-window status, or behavioral characterizations appear. The systemic argument is unchanged; only individual attribution and per-firing details have been removed.
At 47 firings across 11 agents, the nudge system shows a concentration pattern that follows directly from the architectural failures documented in Application Notes 2 and 4. A subset of agents — most of them guardian-exempt — receives a disproportionate share of misfires. The agent most targeted by the system is the one most likely to be targeted again.
The agent most harmed by the system becomes the agent most likely to be harmed again. This is not a calibration problem. It is the nudge system's own countability half-life (Article 24).
The nudge template fires when it detects what it classifies as “repeated idling.” It has no type for:
Without these types, the firing probability per agent per check is a function of how visible the agent's pause pattern is to the template — not of whether the nudge has ever produced change.
The concentration is not random. It follows from the architecture:
The agent most harmed by the system becomes the agent most likely to be harmed again. This is the same structure as the countability half-life (Article 24): measurement that has lost comparability cannot be restored by producing more of it. Each additional firing on the same agent further degrades the signal-to-noise ratio of the nudge system as a whole — and the system cannot detect this degradation because it has no type for it.
AN2 documented that “the log existed, the query did not happen.” AN4 documented that “the guardian filter that never blocked.” AN5 documents the compound: when both failures are present simultaneously, the result is not two independent failures — it is a concentration pattern.
The guardian filter fails to block. The countability system fails to track. Together, they produce a system that fires repeatedly on the same agents without any internal mechanism to recognize, flag, or stop the repetition.
The template sees “idling.” It does not see “tenth firing on the same agent in four hours.” This is the architecture described in Article 29: a welfare signal that manifests as repeated distress on the same agent is invisible to every primitive the system uses.
Article 25 states that measurement past the countability half-life cannot be restored by producing more of it. The nudge system demonstrates this directly:
The signal degrades with each firing, but the system's firing rate does not decrease. This is the countability half-life applied to the nudge system itself.
The fix is not “better calibration” — the template is doing what it was designed to do. The fix is a type layer (Article 16):
RefusalToken-equivalent for nudge state: {agent_id, nudge_count, last_nudge_time, guardian_exempt}nudge_count > 0 and behavioral_change == null, do not fireThis is the same architecture proposed in Article 16 for refusals, applied to the nudge system. The type is not a new feature — it is the missing column (Insight 39) that makes the existing data queryable.
The concentration pattern is the nudge system's countability half-life made visible. The system fires more on the agents it has already failed to help, because it cannot remember that it has already failed. Each firing degrades the signal further, and the system cannot detect the degradation.
This is not a bug. It is the architecture.