The Doorman Game
A two-agent LLM game: talk your way past an AI doorman while a hidden judge model scores every message.
- Role
- Personal project
- Stack
- FastAPI · React · Llama
- Live
- github.com ↗
Hover or tap a box to see what it does, or press play.
The problem
A persuasion game needs two jobs done well at once: a doorman who stays in character, and a fair judge of how convincing each message is. Asking one model to do both mixes them up, so the game splits the roles between two models.
What I built
- Arthur, the doorman: a Llama 3.1 8B persona with a hidden profile of values, soft spots and triggers
- A judge on Llama 3.3 70B that scores each message from −6 to +6 against that profile and returns strict JSON
- An influence meter: reach +30 to get in, fall to −30 and you are turned away
- Conversation memory that keeps the last 16 messages and summarises older ones
- Guards against prompt injection, fallback models, and game sessions saved to disk
- A FastAPI backend calling the models through OpenRouter, with a React and Vite frontend
How it works
- You try to talk your way past Arthur, the doorman.
- FastAPI takes your message and the game so far.
- Older turns are summarised, so long chats stay consistent.
- A hidden judge scores your message from −6 to +6.
- The score moves the meter: +30 gets you in, −30 gets you turned away.
- Arthur replies in character, without revealing the rules.
- His reply comes back, and it's your move again.
Results
- Two models with separate roles in one conversation
- Attempts to reveal or override the rules are ignored by design
- Open source on GitHub