Beyond mediation: an evolutionary benchmark for emotionally and normatively competent AI.

This study presents an agent-based simulation framework for assessing the adaptive and emotional abilities required of AI systems acting as mediators in social and legal environments, where being right does not necessarily mean achieving fairness and stability. Unlike benchmarks based on logical inference or computational efficiency, the framework measures three key traits of effective dispute resolution: emotional regulation, willingness to compromise, and post-conflict trust repair. We defined five canonical negotiation tasks that vary in psychological reactivity, escalation, fatigue in compromise, and interaction-style asymmetry (S1, H1, H2, H3, C1), comparing five mediator policies under identical stochastic conditions using independent Monte Carlo episodes per condition with pre-specified endpoints: agreement probability, conditional time to agreement, inequality of outcomes, robustness across tasks, and a composite score with pre-specified weights that makes efficiency-equity-stability trade-offs explicit. An evolutionary analysis examined long-run policy selection. The results revealed stable trade-offs across scenarios: strong control shortens negotiations yet systematically worsens inequality and does not reliably increase agreement, whereas transparent, weak interventions more consistently balance efficiency and fairness across heterogeneous conditions; slight, reversible framing improves performance primarily in high-reactivity settings. Evolutionary selection favors policies that preserve cross-task robustness rather than optimizing a single metric. The composite objective is intended as a configurable governance parameter rather than a universal social welfare function. Since the baseline calibration intentionally conserves emotional amplitude, we present the current release as an emotion-light v1 benchmark and outline specific v2 enhancements, including event-driven shocks. Empirically calibrated reactivity ranges that make emotion and repair more distinctive. Overall, the benchmark redefines AI evaluation for mediation as process control within the context of emotion and norms, and provides a reproducible protocol for assessing both human and AI mediators.