How do you evaluate the performance of a model in a production environment where ground truth is delayed?
