When models disagree on a prediction, it often signals tricky or risky inputs...

https://paxtonssmartperspective.cavandoragh.org/30-day-plan-to-use-model-disagreement-in-production

When models disagree on a prediction, it often signals tricky or risky inputs worth paying attention to. Measuring ensemble variance or margin helps spot these disagreements. For example, you can route the top 1-2% of high-variance cases to expert review