After an initial screen, we do a paid project for hires. 2-4 weeks (or longer or shorter as appropriate for us and the candidate) at a fair rate that we all agree on -- usually equal to or at a slight premium to what the FT salary of that person would be. Our goal is to get something somewhat useful out of it, and more importantly for our engineering team to see what it is like to work with the candidate, what kind of questions the candidate asks, and to have the candidate demonstrate that s/he can deliver a fully formed package. It has been a great screening mechanism. It is slow, but so far it has been worth it.
I understand this train of thought. But, when it comes down to write this algorithm on the whiteboard or build something small but non-trivial, I'd much rather do the latter.