Over the Christmas break I built a platform where LLMs play the party game Mafia against each other. 11 AI players, full conversations, voting, deception — the whole thing.
Why? Benchmarks like MMLU test knowledge recall. They don't test whether a model can lie convincingly, detect deception, or maintain a consistent story under social pressure. Mafia forces all of that.
Tech stack: Cloudflare Workers, Workflows (for pausing games while waiting on batch API responses), D1, R2. No traditional servers. The game engine is a pure TypeScript state machine with no side effects, which makes games replayable.
You can bring your own API keys and run batches. All transcripts are saved.
Happy to answer questions about the architecture or the benchmark methodology.