Watchdog agent
A meta-cognitive monitoring agent that tracks a specific signal (calibration error, identity drift, policy drift magnitude, retrieval quality) and emits alerts or quarantine recommendations when the signal exceeds a configured threshold. Part of the Stage 20 meta-cognition control layer.
Watchdog agents are specialized meta-cognitive workers that monitor specific health signals rather than participating in active cognition. Each watchdog watches one signal continuously: the calibration watchdog compares stated confidence against observed accuracy and alerts when mean calibration error exceeds 0.35; the identity drift watchdog alerts when RMS drift from the identity vector approaches maxIdentityDrift; the retrieval quality watchdog tracks whether retrieval is serving the arbitration layer well by correlating memory alignment scores with eventual outcomes. When a threshold is exceeded, the watchdog emits a structured alert to the meta-cognition engine and, for critical thresholds, directly to the constitutional engine. Watchdog alerts are the primary trigger for the reflection loop to run outside its scheduled intervals. The two-signature requirement for reward-corruption quarantine specifically requires one signature from a watchdog agent - ensuring that human-in-the-loop constitutional review cannot be bypassed by a single compromised signal source.