Integrating AI into safety-critical systems demands runtime monitoring to avoid silent failures and ensure reliable operation. A single monitor typically only signals an anomaly without giving human supervisors actionable insight. COMET addresses this by anchoring separate monitors in distinct fault modes, merging their outputs in a coordination layer, and mapping the result to a traffic‑light display for human‑oriented feedback.
The framework consists of three monitoring layers: a signal‑integrity layer that catches sensor anomalies; a physics‑consistency layer that checks whether observations obey known physical laws; and an AI competence‑envelope layer that assesses if the model’s prediction lies within its trained confidence bounds. The coordination layer fuses the binary flags from each monitor into a unified status, and the communication layer translates it into red, yellow, or green signals.
COMET was evaluated on an open compressed‑air leak dataset where an AI model detects leaks from acoustic emissions. Test conditions simulated covariate shift, semantic shift, and sensor faults—situations expected at runtime but with unknown exact manifestations. All monitors were frozen after development and tested zero‑shot. The bare AI system failed silently and confidently under all three shift types. COMET reduced the trusted‑wrong‑answer rate to zero on covariate‑shifted no‑leak data and on three of four sensor‑fault cases, while maintaining a nominal false‑alarm rate of about 1%. Single monitors approached the detection performance but could not attribute the fault mode; COMET converts anonymous detection flags into concrete diagnoses.
Residual silent misses under semantic shift define the limits of passive monitoring. Consequently, runtime monitoring frameworks for AI‑enabled safety systems must combine quantitative fault detection and isolation with qualitative, human‑oriented feedback.
Review