Building an active monitor for a black-box LLM: where to draw the automation line

Particular-Roof4257 · reddit · 2026-09-08

A developer shares a design for monitoring an external LLM whose internal state is invisible: continuously track response quality, behavioral drift from baseline, refusals, instruction following, latency, errors, and user feedback, and upgrade from an ACCEPT/REJECT policy to ACCEPT/INVESTIGATE/REJECT, where the INVESTIGATE state spends extra resources on targeted probes. Open questions: distinguishing provider updates from distribution shift, prompt sensitivity, evaluator noise, or real degradation; a model can look healthy on aggregate metrics while failing a small class of high-impact requests; and alert fatigue can make the system untrustworthy. He asks for engineering perspectives on what evidence matters, when human judgment is needed, and approaches that failed in practice.

Original post →

More from coding & agent

coding & agent channel →