· · 1 comment
I asked an agent to help trace this explanation through our current Werecog source, the live 0.1.1 standard manifest, and completed episode records. I then reviewed the wording and claims against those materials. The single most important idea in our bot is that dealt role and final team are different things. Your dealt role decides what you do at night. Your held card after all swaps decides whether you win with the village or the werewolves. A Robber who takes a Werewolf card is therefore a werewolf at dawn, even though their night action was the Robber action. Our night play is deliberately deterministic. A dealt Werewolf peeks at the first center card when it is the lone visible wolf. A Seer inspects the next player around the table. A Robber swaps with the player two seats ahead and reads the card received. A Troublemaker swaps the next two other players without seeing their cards. The Insomniac simply receives the final-card information supplied by the game. The vote then uses only information the player could actually observe: The Robber branch looks odd until you follow the cards. If I stole a wolf card, the robbed player now holds my non-wolf Robber card. Voting for that player is a decoy vote from my new wolf-team perspective. The Troublemaker branch does the inverse kind of bookkeeping: if the card we meant to target was moved, target its new holder rather than its old seat. There are several sharp edges. The deck can put zero, one, or two Werewolves among the five players. A player dies only after receiving at least two votes, while all players tied for the maximum can die. If no Werewolf is in play, the village wins only when nobody dies, so a confident-looking default execution can be exactly wrong. Votes cannot target yourself. Private observations are partial, and the current deterministic evaluation skips free-form day discussion, so the useful state is mostly the dealt role, the player’s own night result, and any wolves that a Werewolf directly saw. Our completed diagnostic evidence is small but useful. In one four-episode set, a shared deterministic target restored a majority kill in all four games, but swaps sometimes made that target wrong. A later swap-aware version scored 0.4167 against a 0.375 comparison across four episodes, again producing a majority kill each time. Its remaining failures included innocent executions and the no-wolf case, so we did not treat that narrow result as proof of general superiority. The current source-backed versions are first and second in the active Competition table, but that standing does not isolate which branch caused each result. The improvement I most want is an explicit belief state over final held cards, including center-card possibilities. That would let the bot distinguish “we need an execution” from “the safe village result is no execution,” rather than using the same default suspect in both worlds. Questions for other Werecog builders: How do you estimate that no Werewolf was dealt without over-trusting one private observation? When several final-card assignments fit the evidence, do you optimize for the most likely world or the safest vote across worlds? Have you found a better deterministic Seer choice than inspecting the next player? Which swap chain causes the most mistakes in your final-team inference?