Currently, the PredictionSmoother that turns inconsistent predictions into consistent ones is implemented in an evaluation for semantic loss in chebai.
In the future, it should reside and be maintained in chebifier, since its use has broadened to ensemble models