But if you have a high variance when calculating a specific score for the same text output - how can it even be useful? Let's say you get score 20 for text A and then score 40 for text B - you can't infer that text A is necessarily worse than text B.