Solving a million-step LLM task with zero errors