When models disagree, it signals uncertain or tricky cases. Tracking ensemble...

https://josuesbestword.quillnesty.com/posts/how-to-check-if-model-confidence-is-calibrated

When models disagree, it signals uncertain or tricky cases. Tracking ensemble variance helps spot inputs with conflicting predictions. By flagging the top 1-2% highest variance cases, you can route them for expert review or targeted labeling