Unsafe Consensus in Diagnostic Deliberation

Multi-agent deliberation can improve high-stakes classification, but agreement is not a safety certificate. We study diagnostic-style decisions in which agents with asymmetric expertise may converge before high-cost alternatives have been excluded, a failure mode we call premature diagnostic closure. We formulate deliberation as a protocol-control problem without online ground truth: a mediator observes agents' reports, optional justifications, calibration confusion matrices, and an asymmetric loss, and must decide whether to continue, challenge, certify, or escalate. We propose Diagnostic Consensus Energy Minimization (D-CEM), a loss-aware controller that uses each agent's confusion matrix as a map of plausible differential diagnoses. Given current reports, D-CEM forms confusion-induced posteriors, identifies the most dangerous plausible miss under the loss, and computes a diagnostic consensus energy combining posterior disagreement, expected harm, and loss-aware margin. The resulting policy continues while risk decreases, issues targeted differential challenges, certifies only low-energy large-margin decisions, and escalates when high-risk deliberation stagnates. We prove loss-sensitive certification and cost-aware stopping guarantees. Across synthetic diagnostic tasks and clinical LLM datasets, D-CEM reduces expected diagnostic cost, high-risk misses, and harmful consensus while controlling communication.