An interesting idea, though there are a few complications.
The version where the voter selects which receipt will be displayed before leaving the site does a nice job of handling external coercion, but it also means that the system knows which voters will never be able to verify their vote *before* the votes are counted. If it isn't possible to decide to check any individual vote after the election, then there's still an opportunity for tampering. There are steps you could take to make tampering less likely - say, recording the "use A/B" choice and the "display A/B" choice on separate hardware and combining them only after the vote count - but none are foolproof against someone with access to all the hardware and software being used.
Second, and to my mind, most important - as a voter, I don't particularly care whether my individual vote is counted correctly. In elections where outcomes are often decided by a few percent of votes, an absolute minimum requirement is that every single vote *could* be verified. But, to really be effective, a system really ought to insure that a significant fraction of votes will be verified. If only a small fraction of people do check their receipts - and, if someone is sufficiently clever about incorporating demographic information into vote-rigging and is willing to accept a small but non-zero number of accusations of receipt errors - then fraud is still quite viable.
Not a bad idea to play with, but on the whole it sounds like a very complicated fix, compared to something much simpler and likely cheaper, such as a paper receipt that every voter can look at in real time and which then gets placed into a lock-box which can be independently observed with as much scrutiny as we're willing to pay for.
