To me the challenge is slightly different. With online shopping, you get hundreds, thousands, millions of tries to get it right. If someone’s new socks dont turn up, learn why not and try again i.e. fail fast. How many tries do you get with online voting given it happens once every one to four years? So ideally, it needs to start somewhere even less relevant – school elections, clubs, professional bodies, etc. and slowly work its way up. But even then it’s unlikely that you’ll find all the problems until you go for it at scale, at which point it’s too hard to try again.
As for A/B testing – why not get people to fill in a paper ballot and an electronic one at the same time? Or use the electronic machine to print a “receipt” (i.e. a completed paper ballot) which is then counted as a paper one would be?
finally, what about other things like objections e.g. actively spoiled ballots, consciously submitted blank papers etc?