The model's internal confidence score for each word it generates, derived from its output distribution.