Identical models don't imply identical preferences or emotions. Our brains can differ a lot even if we predict the same stuff.

Yes, but the two will have identical maps of their own preferences, if I'm understanding your scenario. They might not in fact have the same preferences, but they'll believe that they do. Brains and minds are parts of the world.

Hm, they sure do to me

Based on what you're going for, I suspect the right heuristic is not 'does it convey information about an attitude?', but rather one of these:

• Is its connotation more important and relevant than its denotation?
• Does it purely convey factual content by implicature rather than by explicit assertion?
• Does it have reasonably well-defined truth-conditions?
• Is it saturated, i.e., has its meaning been fully specified or considered, with no 'gaps'?

If I say "I'm very angry with you," that's an empirical claim, just as much as any claim about planetary orbits or cichlid ecology. I can be mistaken about being angry; I can be mistaken about the cause for my anger; I can be mistaken about the nature of anger itself. And although I'm presumably trying to change someone's behavior if I've told him I'm angry with him, that's not an adequate criterion for 'empiricalness,' since we try to change people's behavior with purely factual statements all the time.

I agree with your suggestion that in disagreements over matters of fact, relatively 'impersonal' claims are useful. Don't restrict your language too much, though; rationalists win, and winning requires that you use rhetoric and honest emotional appeals. I think the idea that normative or attitudinal claims are bad is certainly unreasonable, at least as unreasonable as being squicked out by interrogatives, imperatives, or interjections because they aren't truth-apt. Most human communication is not, and never has been, and never will be, truth-functional.