How would you evaluate whether a new model is better than the current baseline beyond accuracy alone?
