Artificial intelligence is entering a phase in which its capabilities increasingly exceed the mechanisms available for oversight and control. As systems become more autonomous, adaptive, and integrated into critical infrastructures, the need for external, real-time governance becomes essential. This work presents the first operational implementation of the Sentinel System, an external AI oversight architecture designed to monitor, evaluate, and regulate model behavior in real time, independently from the models it supervises. The prototype introduces a structured decision layer capable of analyzing proposed actions, assigning risk levels, and enforcing outcomes including allow, modify, and block. Unlike traditional approaches based on isolated step-wise evaluation, this system incorporates trajectory-aware analysis, enabling the detection of behavioral evolution across sequences of interaction. Through a controlled experimental framework, the system is evaluated across both isolated inputs and multi-step behavioral sequences. Results show that while step-wise evaluation can classify individual actions correctly, it fails to capture progressive escalation patterns. The trajectory-aware extension allows the system to detect early deviations and intervene before critical thresholds are reached. The paper introduces a minimal formalization of this approach, including risk aggregation across time and a decision policy based on cumulative behavioral context. These elements provide the foundation for a new class of AI oversight systems capable of interpreting behavior as a continuous process rather than discrete events. This work represents the first step toward a scalable, verifiable, and operational infrastructure for AI governance, establishing trajectory-aware oversight as a necessary principle for the safe development of advanced artificial intelligence systems.
Emanuele Colombo (Tue,) studied this question.