The Bactra Review   Individual Strategy and Social Structure
In particular, his models show ergodicity. Basically, this means that if you have a large number of such systems, and at any one time p percent of them are in a given state, then every single one of them will spend about p percent of its time in that state. (Technically, there's a distribution over the states such that the state-space average of any well-behaved function equals its time-average starting from almost any state, if you take the latter over a long enough interval.) Ergodicity is a very strong property, and one of the stronger predictions his models make; it goes away if the agents are not perverse, or can remember unlimited stretches of the past, in which case each population will lock in to a single convention forever. It is also liable to be hard to test empirically, because it can take a considerable time to converge on the ergodic distribution, and in the meanwhile the game people are playing is apt to have changed.