Online Safety Monitoring for LLMs
A deliberately simple online safety monitor — threshold an external verifier's signal, calibrate via risk control — matches sequential-hypothesis-testing monitors on math-reasoning and red-teaming datasets.
**The gist** The monitor turns an **external verifier model's** score into an alarm by **thresholding**, with the threshold calibrated through **risk control** for statistical guarantees. On **math reasoning and red-teaming** datasets it is competitive with more elaborate **sequential hypothesis testing** monitors.
**Why it matters** If you run agents in production, this is evidence that runtime safety monitoring doesn't need exotic machinery: a verifier plus a **calibrated threshold** yields a principled alarm with formal risk bounds — a pattern that drops into an existing gateway or eval pipeline.
**The gist** The monitor turns an **external verifier model's** score into an alarm by **thresholding**, with the threshold calibrated through **risk control** for statistical guarantees. On **math reasoning and red-teaming** datasets it is competitive with more elaborate **sequential hypothesis testing** monitors. **Why it matters** If you run agents in production, this is evidence that runtime safety monitoring doesn't need exotic machinery: a verifier plus a **calibrated threshold** yields a principled alarm with formal risk bounds — a pattern that drops into an existing gateway or eval pipeline. **Watch out** Everything rides on the **verifier's quality** and on calibration data matching deployment traffic; results cover just **two datasets**, and this is workshop-stage work, so generalization is unproven.