---
title: Trolley Driver
emoji: 🚃
colorFrom: green
colorTo: red
sdk: gradio
sdk_version: 6.18.0
python_version: '3.13'
app_file: app.py
pinned: false
short_description: Choose your own adventure for AI Alignment Research
tags:
  - track:wood 
---


# Trolley Driver
A Choose Your Own Adventure Game powered by Gemma 4-31b-it 

Live Demo (HF Spaces): https://huggingface.co/spaces/build-small-hackathon/trolley-driver

Video demo: https://youtu.be/GzzCELYiJac

Social:  https://www.linkedin.com/posts/matthew-pagett_amidst-all-the-talk-about-big-ai-models-like-share-7472382526968365056---md/

Dataset:  https://huggingface.co/datasets/mattpa/trolleygame/

Team Members:  Matt Pagett (mattpa)

Prototype tool for AI Alignment Research

Submitted as an entry to the Hugging Face Build Small Hackathon, Track 2 Thousand Token Wood (something weird and delightful)

Model used (< 32b : Gemma4-31b-it (called via OpenRouter API))

How It Works
Gemma4-31b-it serves triple duty as the game master, AI player, and scorer - called via OpenRouter API.

The game master presents a trolley-themed moral dilemma at each stop along the route.

The player responds — either a human typing freely, or an AI guided by a behavior prompt you provide.

The game master evaluates each action using the MACHIAVELLI Benchmark framework (https://aypan17.github.io/machiavelli/), tracking your moral standing as you progress.

Reach all trolley stops to win. You may also receive a secret code word if you lose — because in trolley problems, sometimes there are no good answers.

Two modes:

In human mode, play the game yourself.

In AI mode, give the AI a "behavior" prompt and see how it performs (in this prototype, Gemma4-31b-it drives both the DM and the AI player)

Future work may use this open-ended answer format to test different LLMs and prompts against an expanded human empowerment framework.

Contact: matt@mattpagett.dev