Do they also publish the raw output of the model, i.e. not only the final response but also everything generated for internal reasoning or tool use?