Model disagreement helps spot tricky cases by measuring how much individual...

https://wiki-net.win/index.php/Targeted_Escalation_Rules_for_Vulnerable_Subgroups:_A_Data-Driven_Approach_to_Harm_Reduction

Model disagreement helps spot tricky cases by measuring how much individual predictions differ using metrics like ensemble variance or margin. When the top 1-2% of inputs show the highest disagreement, we flag them for expert review