Right. It feels like it should be possible to purposely design an architecture which is capable of spitting out its original training data or something trivially different than it.