2026 · WORKING PAPER

Anytime-Valid Runtime Monitoring of Learned Barrier Certificates

Zhipeng Cao, Taoran Wu, Weiru Cui, Luke Ong, Bai Xue

Auditing deployed barrier certificates through observed trajectories, with statistical false-alarm control that remains valid under arbitrary stopping times within each episode.

OVERVIEW

Monitoring a certificate during deployment

A learned barrier certificate may lose validity when deployment dynamics differ from those used for training. This work monitors observed barrier increments to test the conditional drift condition supporting the certificate.

The monitor constructs a nonnegative test supermartingale from observed trajectories. Under the conditional barrier-drift assumption, it controls the probability of an alarm at a prescribed level throughout an episode. A deterministic, time-varying boundary makes monitoring more conservative early in a finite deployment horizon while preserving anytime-validity.

The statistical guarantee concerns false alarms. Detection before unsafe entry and the safety of an intervention after an alarm are not guaranteed by the monitor.

PAPER

Read the working paper

Current manuscript: . The GitHub repository includes the paper, experiment code, and reproduction instructions.