2 ms·

I agree that 82% to 4% seems pretty robust. However, larger data sets are more robust to potential flaws in experimental implementation.