Interposes between an LLM policy compiler and a networked agent population to balance cooperation gains against manipulation risks, using hard constraint filtering plus soft penalized-utility optimization evaluated by an Ethical Cooperation Score (ECS). Tested on scale-free networks with 80 agents under 70% adversarial conditions, achieved 14.9% ECS improvement over unconstrained optimization (0.741 vs 0.645) while reducing hub-periphery exposure disparities by over 60%. Accepted at AMSTA 2026.