The language model layer: what it reads, what it may say, what it may never say, and how each answer is hashed.