Most expert work isn’t “produce a probable artifact”; it's "choose a good move considering other agents, guessing hidden state". LLMs default to single-shot artifacts and need World Models to progress