So, the other day I played an interesting 15 minute game on Instant Chess against a fairly decent player (rated mid-1800's) and afterward went over it with Stockfish 6 and Komodo 8. The position was closed with my opponent's pieces all crowded together on the back three ranks and I made a couple of offers of a Pawn; that's the general scenario. The result was similar to the Yermolinsky game in that sometimes the evaluations didn't always match up with the Shootout results; same for my gut feelings. Sometimes my position looked overwhelming and the evaluations confirmed it, but Shootouts showed something unexpected. Another thing that happened is that in a couple of positions the engine's evaluations drastically changed AFTER its recommended move was played.
Things like this happen a lot and when they do, it's confusing unless you are rated up around 2600. That's why I always advise believing the GM when it comes to positional evaluations. It's also why experts advise that engines need opening books. The problem is, when analyzing a game like the one I played, there's no 2600 rated GM sitting here with me that can offer any insights when the engine evaluations are not clear.