One I was just working on involves picking two random points, and if they are in the wrong order, reversing the entire subrange between them. I found the algorithm on http://richardhartersworld.com/cri_d/cri/2001/badsort.html, which says that the average case is is probably somewhere around O(n^3) or O(n^2 log n) (he's not really sure).
I think it might be possible to do it more efficiently, because I think it might be possible to do the reversal operation in O(1) time.
Actually, I just realized that doing that would make the whole thing I say maybe because I just realized that the data structure I had in mind would put accessing the random elements at O(log n) and determining if it needs reversing at O(n).