At the moment, only the case where an InProcess model fails is handled. This is done by returning a NaN Prediction object. It's not immediately clear to me what a failed pre-process method (or even a post-process method) should return. Maybe the original input/preds?