It's often claimed the models will perform best in the harness they're trained in but all the train/eval time incidents/cheating we see that don't seem to occur in normal use is maybe a clue the opposite may be true. You need to let the model feel this is real life.