Maybe the trick is to hide vetted correct code, of whatever origin, behind function calls for documented functions, thereby iteratively simplifying the work a later-trained LLM would need to do?