Disagreement Is All You Need: Adversarial Multi-Agent Deliberation with Three-Loop Reinforcement Learning | Synapse