Simulated Social Deliberation: A Multi-Agent Framework for Studying AI Bias, Identity Leakage, and Epistemic Inequality in Structured Debates

We present Omnibrain, a multi-agent deliberation framework for simulating structured debates between LLM agents instantiated with distinct social identities. Across three single-run phases — expert panel (unanimous approval), identified social profiles (7/3 rejection), and anonymous mode with hidden fact-checker (9/1 approval) — we observe eight candidate patterns requiring replication: partially prompt-influenced identity leakage through vocabulary despite anonymization, in-group moderator bias, higher actionability of non-expert profiles, spontaneous convergence on physical presence as a right, epistemic regression co-occurring with vulnerable testimonies, self-correction as an integrity signal, and structural exclusion reproduced by the debate design itself. These are preliminary observations from an ongoing research program, not confirmed causal relationships. We describe the framework for replication and outline planned ablation studies and cross-model comparisons.