Of those 21% how many are time-dependent questions that are past the model’s training and requires research to verify? Like the “did Ukraine attack Russian in the past week” question?