AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-21
Anytime-Valid Runtime Monitoring of Learned Barrier Certificates
Learned barrier certificates for stochastic systems can lose validity during deployment when the underlying dynamics differ from training. The authors propose an anytime-valid runtime monitor that audits the certificate from observed closed-loop trajectories, formulating barrier drift as an episode-wise sequential hypothesis test whose increments form a nonnegative test supermartingale, giving finite-sample false-alarm control under arbitrary stopping times. The monitor detects statistical evidence that the certificate no longer matches the observed dynamics—rather than predicting unsafe entry—and its alarm behavior was evaluated under valid and invalid certificates.