"users": 
        {
         "first":["Homer","Hank","Peter"], 
         "last":["Simpson","Hill","Griffin"]
         },

saving a few characters in the process. though I see that the advantage of this system is supposedly that it can handle any sort of shape of data, not just ones with a fixed schema. I've been trying to figure out how trang [http://www.thaiopensource.com/relaxng/trang.html] does its schema inference trick. (it turns a set of xml files into a relaxNG schema). If you have a schema for a JSON file, it's knowledge you can apply to algorithmically creating really efficient transformations.