This is a great idea, but it doesn't "force correct code" any more than unit tests do when applied to human-written code. It also can't force the code to have other desirable properties such as readability (granted that this doesn't usually seem to be a problem with LLM output).
