There's a thread in scripting started by me where he goes into a bit of detail about it.

If I interpret it correctly, he's simply creating a detailed level in WED (??), complete with shaders, or model based given the shaders. He sets his camera positions, and takes a snapshot of the level with everything turned on from the camera. This is his background, and his hidden scene geometry is in place and should match. I would suggest. XxX that you simplify the hidden geometry if non-visible geometry has an influence on the framerate. There's no reason to use the super detailed version for simple collision. Per Wintermute's techniques, the only parts of the collision hull that require extra detail are the portions that stencil shadows should cast on. Everything else can be simple cubes or whatever.

For go behind/go in front of object, he'll need to take snapshots of those isolated sections and place them via whatever method he chooses. Some kind of triggering method will need to be established if he uses panels and layers. If the player falls below a certain horizontal value then the player can walk in front of these objects and if they're above it then it can go behind it. I think Wintermute does this based on the origin which is generally located at the bottom center of the isolated sprite/panel element.

You can read all about how they do it in their latest manual:

http://www.mediafire.com/?em93txnmg3q