[atrifix]

Okay, this is a possible way of solving the paradox. After all, we can reject any of the assumptions, and it seems like in real-life situations information is definitely partial. I don't want to hijack this thread, but I'd argue that there are certain situations where assumption #5 applies that you're still going to want to maintain that it's rational for people to cooperate, so one of the other assumptions must go as well. Consider this quasi-centipede game: on each round, we play a simultaneous-move prisoner's dilemma. Defecting when the other player cooperates pays (5,0), both defecting pays (1,1), and both cooperating adds 3 to each player's payoffs and keeps the game going another round. The game lasts for a finite number of rounds t (say t=3000). Now, if both players cooperate every round, their payoffs are (3004,3004), but in spite of this, there is a unique equilibrium where both players defect on the first round and get (1,1). I suppose that we could maintain that defecting on the first round is the rational play, but that seems pretty counterintuitive.