A sensor transmits. A sentinel halts.
Most "AI watchdog" mechanisms are sensors — they log, they alert, they generate reports. What they cannot do is stop the system. Without halt authority, oversight becomes a downstream theatre 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.