Model disagreement helps spot risky inputs by measuring uncertainty with...

https://magic-wiki.win/index.php/Disagreement_as_a_Diagnostic_Metric,_Not_Noise:_What_Does_That_Mean%3F

Model disagreement helps spot risky inputs by measuring uncertainty with metrics like entropy or ensemble variance. When models disagree sharply, those top 1-2% cases get flagged for human review or targeted labeling