making models target benchmark about being good at lying and getting away with it (werewolf) is certainly an interesting choice