The biggest issue (aside from computational cost) is definitely how to reconcile conflicting models, although no one would ever be editing the entire model, only small parts of it. I hope (and I could be wrong) that once the system reaches a certain critical mass, predicting the emergent behaviour from the microscopic details becomes so hard that someone with a political agenda couldn't easily come up with ways to manipulate the system just by making a few local changes (you can think of this as similar to a cryptographic hashing problem). Other large-scale systems (like cryptocurrencies) derive security from similar 'strength in numbers' principles.

One option is to limit input to the system to only peer-reviewed statistical studies. But this isn't a perfect solution, for various reasons.

Using a connection to prediction markets (so that people have some skin in the game) is a nice idea, but I'm not sure how you're thinking of implementing that?