A sensor transmits. A sentinel halts.
Most "AI watchdog" mechanisms are sensors. They log, alert, and generate reports. What they cannot do is stop the system. Without halt authority, oversight becomes a downstream theater performed on frozen evidence.
The Grievance Watchdog Architecture resolves the paradox by giving the sentinel negative power only: the capacity to block continuation under narrowly scoped triggers, with mandatory human re-entry and attributable overrides. It cannot act; it can only refuse. That refusal keeps the affected party reachable before the record closes.