So an LLM (or more realistically, a huge swarm of agents) should check his work, find a mistake or gap, or, if there is none, provide a formal verification.