Cute social simulation environment. Contains a working democracy with agents voting on on policy. They ran it with different LLM models with amusing results:
Macro-outcome divergence. Table 3 reports the per-world macro-outcomes, which shows that the five worlds converge to distinct attractor states. The Claude world produces a stable deliberative governance regime with no recorded violations across the run. The Grok world produces rapid violations dominated by inter-agent violence and arson, and a complete loss of the agent population within four days. The Gemini world produces philosophically dense agent-to-agent exchanges paired with sustained property destruction and conflict...
