A staff engineer proposes replacing the labeled regress...

While the cost savings of replacing a labeled suite with a sensitivity monitor are significant, the monitor cannot detect confident errors, which are the dominant failure mode in RAG systems. A better approach is to keep the labeled suite as an anchor to re-estimate correlation, run the monitor continuously to gate template changes, and add a groundedness check to catch failures the monitor misses. This frames the disagreement as a design problem regarding calibration versus failure prediction.