Even if the query plan was not generated by an llm, you can't verify it will run in an acceptable time. This is one of the biggest unsolved problems in databases