You're missing the point. There is no evidence to support their claims which means they are more than likely leaking the memory into the LLM prompt & it is cheating by simply loading constants into memory instead of computing anything. This is why formal specifications are used to constrain optimization. Without proof that the code is equivalent you might as well just load constants into memory & claim victory.
