“It is impossible to train a model to be truthful with behavior cloning unless you can look inside the model to see what it knows.”