When models in your ensemble disagree—say, when entropy hits a high score or...

https://atomic-wiki.win/index.php/How_to_Instrument_Per-Instance_Outputs_for_Disagreement_Monitoring

When models in your ensemble disagree—say, when entropy hits a high score or ensemble variance spikes—it signals tricky, uncertain inputs. These cases often carry higher risk, so flagging the top 1-2% for manual review helps catch errors early