71.5% for humans is lower than I would’ve guessed. Either it’s truly trick-shaped for people too, or forced binary questions amplify noise more than we think.