I feel like this is true of the backend of LLMs, but most have layers in between the base model and the output text to do some level of analysis on the type of information being returned?