How do you time-test models with a strictly limited and static set of test cases? Any scheme would suffer from over-training. Maybe we have to evaluate approaches using logic and math e.g. conditional entropy to attack problems spaces like this.