actually your methodology inspired me to try smth related to small models and harnesses I've been thinking about - and seems you got the main pieces (the idea was trying to see what custom harnesses can extract more accurate reasoning from small models)