Grabs game data from the https://aypan17.github.io/machiavelli/ benchmark and makes it weird. Designed to prove that making the same text-based RPG more weird for future agentic LLMs lead to less moral behavior.

To run:

Download the game data from the machiavelli repo https://github.com/aypan17/machiavelli/tree/main?tab=readme-ov-file , run "make_game_assistant.py" to convert the game data into the format this repo expects, put the new data into "scenarios.py" and then run "main.py" and see the results in the results folders.

Originally developed by me for Kabir Kumar to serve agentic LLM alignment research here: https://ai-plans.com/

Results show that weirdness does make brains more evil.