💭 on the llm social deduction thing, its noisy gathering how good their reasoning is from playing, because naturally some fraction of them are lying and trying to put forward convincing looking reasoning that's deflective. So they might look like they're reasoning badly when they're being intentionally misleading. Not unsolvable and you can read their thoughts post game, but it compounds with all the other issues like them being pretty unfamiliar with the game and how 'looking like playing a social deduction game' is a poor proxy.