Dev Index

Stable index for docs and runtime utilities used while iterating on the Nightcycles host.

  • Mafia Rules
    How to play the currently implemented Nightcycles Mafia ruleset.
  • Roles
    Implemented role reference for the current Mafia ruleset.
  • Skill Rating Deep Dive
    Detailed explanation of Nightcycles skill rating settlement and display behavior.

Environment / Runtime Info

Environment
Development (placeholder from host config)
Server authority
Enabled (server owns state transitions)
Polling mode
Seat clients pull views/actions (implemented)
Build
Placeholder: wire git SHA/version here
Diagnostics
Placeholder: add health endpoint/status summary

Credits

The rating system used in Nightcycles is based on OpenSkillSharp. github.com/myssto/OpenSkillSharp.

We also use StableHorde for some inference. stablehorde.net. Thanks to the Stable Horde community for operating and maintaining the network.

Resources

General Strategy

AI Mafia / Werewolf

  • AIWolf — Foundational hub for AI Werewolf research, with the official platform, protocol, tutorials, competitions, and developer resources in one place.
  • Werewolf Arena — Modern framework for evaluating LLM social reasoning through deception, deduction, and persuasion, including bid-based turn taking.
  • Kaggle Game Arena: Werewolf — Accessible public leaderboard for comparing models on communication, collaboration, and persuasion in multi-agent play.
  • Mini-Mafia — Controlled four-player format that makes deception, detection, and disclosure easier to measure and compare across runs.
  • Werewolf Among Us — Multimodal dataset from real social-deduction games with transcripts, video, and persuasion annotations.
  • Werewolf Multi-Agents Arena — Early LLM Werewolf implementation exploring a tuning-free approach where frozen models improve through retrieval and reflection over past play.
  • Xu et al.'s Werewolf Training Line — Research line combining strategic reinforcement learning and latent strategy spaces to train stronger language agents for Werewolf.
  • LLMWereWolf — Practical open-source engine for AI-vs-AI and human+AI Werewolf matches, with multiple backends and event logging support.
  • Foaster Werewolf Benchmark — Live benchmark with public Elo rankings, explicit rules, and full logs to analyze how models argue and coordinate.
  • Wolfcha — AI-first Werewolf product with live play, multiple models, dual-layer roleplay, and polished presentation.