[root_axis]

Because the model can output data in a manner optimized for training a new model, including outputs that were post-trained like RLHF and RLVR.