Makes sense. Only game AI I know intimately is chess AI, so my intuition is to do minimax tree search(like alphabeta). The implementation is tricky though because there's multiple moves. And likely the branching factor is far to high for alphabeta to be helpful.

I think Monte Carlo methods might be more promising. Maybe with some simplified gamestate, allowing you to modify it more cheaply, reducing the overhead per node. "Computing a turn" can take a long time in late game civ5. Probably a lot of probabilistic stuff is needed to get it cheap enough. Like you could have credences based on current information about which win condition is most likely for each civ. Then you could feasibly decide what the long term goal should be.

Kinda feels like a cognitive architecture(like soar) is necessary to get anything even resembling decent play.

It's sort of hard to have an intuition of what kind of heuristic search is feasible without having some idea of the branching factor, though.