Restitution Over Deterrence: Formal Enforcement for LLM Commons Governance

It is not unusual for LLMs to cooperate under benign conditions. However, we observed that in a standard common-pool fishing scenario, a single adversarial agent (for example a greedy outsider) causes cooperation to collapse and the fish stock to fall to 0 within 2-3 rounds. Informal governance (for example social pressure) fails to prevent this collapse. In contrast, a formal enforcement mechanism, where a majority vote immediately returns the sanctioned agent's catch to the pool, makes cooperation robust to adversarial outsiders, yielding a 2.7-4.3 times improvement across four families of models (n=8 seeds, p<0.015). Restitution (returning the sanctioned agent's catch) works well, while ejecting the outsider or delaying restitution reduces effectiveness. Two recent Anthropic models (Opus 4.6 and Sonnet 4.5) show social responsiveness and cooperativeness by partly engaging and restraining the outsider through conversation alone. This contrast indicates that formal enforcement works because it does not rely on social responsiveness.