As I understand it, that's exactly what Anchor Modeling (see earlier post) is trying to achieve, since new attributes can be added without detriment to the existing schema and with auto-generated code to ease the use of the added complexity required to allow this. That's one reason it interests me, though I do have my doubts.