I'm sure it's no where near good enough yet, but a legal model getting the answer right 83% of the time is still quite impressive imo.