It's unlikely that a large volume of text will be generated at the base rate (errors compound), so the number might be higher than we expect.