If the training data contained censorship related prompts, any transfer could simply reflect the student directly learning the behavior. Only distilling on finance tasks and separately evaluating on political censorship tests if the teacher's censorship behavior transfers through unrelated outputs at large model sizes, i.e. subliminal learning (https://arxiv.org/abs/2507.14805).