I agree, but also via some napkin math, the chance of getting those different results is something like 40%.

(Ie, if you sample the same signal twice for the numbers they did in the study, there’s a 40% chance it’ll be >0.04 away from the original sample, as numeracy was from language)