When models disagree, it's a strong signal something's tricky or risky about...

https://wiki-triod.win/index.php/How_to_Stop_Over-Flagging_Disagreements_That_Are_Just_Noise

When models disagree, it's a strong signal something's tricky or risky about the input. By tracking metrics like ensemble variance or margin, you can spot the top 1-2% most disputed cases