AI Council framework uses architectural heterogeneity to preserve disagreement in LLM policy simulation

The AI Council is a three-phase multi-agent deliberation framework that assigns different 7–9B parameter LLMs to distinct value perspectives, significantly reducing artificial consensus in policy simulation. Experiments across 120 deliberations show architectural heterogeneity outperforms single-model systems at preserving substantive disagreement.