AI Model Showdown

Problem

Turns multi-model conversations into an observable game. The goal is not to get one final answer, but to expose how models disguise intent, challenge each other, vote, and leave subtle wording traces. It is useful for demoing multi-agent dialogue, model persona differences, and lightweight rule-based interaction.

How It Works

• God-view setup: randomly selects a daily topic and assigns one wolf among four roles, then reveals the wolf identity and both anchor sets to the user.
• Betting loop: the user only chooses good wins or wolf wins and does not participate in speaking or voting.
• Two-round flow: roles make statements, suspicions, and votes; the first round has four players, the second has three, and eliminated roles stay silent.
• Retrospective: outputs identity reveal, clue replay, vote replay, and difficulty rating to review wording drift.

Boundaries

This is a casual conversational game with everyday topics, not a persistent leaderboard or production task system. On multi-agent platforms it can present different models; otherwise the current model simulates all roles. The design favors short lines and a fixed game surface, so it is not suitable for serious evaluation or production-grade orchestration.