You're grasping for a reliable unsupervised truth machine. That's a fundamentally intractable problem until you limit it down to a wolframalpha clone. And not even that by LLMs.