When models disagree on a prediction, metrics like ensemble variance or entropy...

https://milosinterestingcolumns.huicopper.com/how-to-build-a-conservative-rule-when-disagreement-is-moderate

When models disagree on a prediction, metrics like ensemble variance or entropy can highlight risky inputs that need attention. For example, you can track the top 1-2% of cases with the highest disagreement and route them for manual review