To play devil's advocate, a lie detector isn't really supposed to measure meaningless lies. It's about emotion and nervousness and the like. Math questions about a randomly assigned number would likely not generate much of a meaningful signal, and thus this test would not be expected to work.
It sounds like you're arguing that a lie detector only detects lies that the subject considers "meaningful".
So now you have two problems: Detecting whether a subject believes a particular topic is "meaningful", and then detecting whether they are lying about it.
I think interrogators are missing the point: If your interrogation technique can't even produce information that is verifiably correct, then why on earth would you believe any information produced that is non-verifiable?
