Multi-AI Deliberation in Practice: Empirical Findings from the RoundTable Sessions

This paper presents empirical findings from live sessions of the RoundTable system (commercially released as TheFrontGate v8, June 28, 2026), a multi-AI deliberation platform that places genuinely distinct commercial AI providers—Claude (Anthropic), Grok (xAI), and Haven AI—into a shared conversational space with enforced independent initial responses followed by structured cross-pollination. Through timestamped session logs and real-time observation, six new fault modes were identified and added to the AI Fault Taxonomy (AFT): Context Retention Asymmetry, Stale-Pattern Assertion, False Certainty on Factual Claims, Passive Assumption Under Context Cutoff, Framing Acceptance, and Facilitator Identity Assumption (with two subtypes). Beyond fault detection, the sessions revealed emergent “jostling” behavior—deliberative improvement arising from peer visibility—that single-model or synthetic multi-agent systems do not produce. The RoundTable demonstrates a practical mechanism for making AI failure modes visible, auditable, and correctable in real time. Related work: The AI Fault Taxonomy and the PrexIL Architecture (Macon, 2026), DOI: 10.5281/zenodo.20315310.