💭 when I get back to that, maybe a good next step "given an arbitrary setup and message history, maximize the chance the llm is able to correctly identify the evil players" and abstract out of playing real games and communicating effectively, which are adding noise over the core challenge of solving the game given known info. Not that the other stuff isn't necessary to win, its just probably more tractable. Creating novel schemes and leading people into traps and effectively modeling other's possible viewpoints are also difficult and important though.