I've long felt the key is to get the interviewee to do some real work for a day. Pair programming seems like it might work well.

You'd have to come up with some suitable toy problems that you could make significant progress on in a day.