My first thought is that one of the big departures Dragon Box has, is that it removed the central concept (math), and just kept the patterns (balancing equations).

Second, the problem with probabilities is that if the best you can do is 90% within the rules, then you'll be failing 10% of the time, and people don't like that.

The second problem seems easily solved: come up with a situation where you run the experiment N times and reap the rewards of your strategy. Prisoner's Dilemma tournaments, for example. Given we're dealing with probabilities, ANY sort of gambling, and just play out the results based on statistical odds (or, at higher levels, larger samples. You could even have the elite level be single-instances :))

The game still focuses on gathering information and learning about the system, and it probably needs to have a fairly concrete UI and mechanic for measuring and representing that. But it removes the social and cultural aspecs, and focuses it on a very simple, easy to understand system.

I'm not sure as to specifics, but a game where the goal is simply to learn how to play the game seems interesting, and then there's a meta-game that is the UI and mechanics that go in to teaching you and ensuring this isn't a ridiculously frustrating black box to deal with :)

Having multiple possible rules and permutations is important, of course. An unchanging tutorial is probably a good place to start, though. If you can write that, and it's interesting and works, you're probably on to something useful :)