To me the main misstep with a lot of modern interview testing is the time-constraint factor. Having 20 minutes to solve a complex problem sans-Google in a not-really-a-text-editor field on a web page just isn’t a real-world scenario. And, what you do in that setting doesn’t really say anything useful.

If you’re going to test people, either whiteboard it and make it absolutely clear you just want to see how they break it down; or, give them hours to do it right without a lot of restrictions.