Respecting evidence: belief functions not imprecise probabilities