I agree that this isn't a very interesting example, but your statement is: "just asking the model to do a simple transform". If you assert that it understand when you ask it things like that, how could anything it produces not fall under the "already in the model" umbrella?