When models disagree strongly—say high ensemble variance or low margin—it flags...

https://station-wiki.win/index.php/What_Is_Selective_Prediction_and_How_Does_It_Reduce_Harm%3F

When models disagree strongly—say high ensemble variance or low margin—it flags inputs where predictions aren’t stable. These risky cases help teams focus efforts by routing the top 1-2% of disputed inputs for human review