From Situational to Unconditional: The Spectrum of Moral Commitment Required for Multi-Agent Survival in Non-linear Social Dilemmas

We establish the Moral Commitment Spectrum: a systematic relationship between environmental severity and the minimum moral commitment required for multi-agent system survival. In linear PGG environments, situational commitment achieves group-level ESS. In non-linear environments with catastrophic tipping points, pure RL fails (Nash Trap at λ≈0.5, 5.3% survival), and only unconditional commitment (φ1*=1.0) guarantees survival. Key findings: Decentralized baselines (Inequity Aversion, Social Influence) achieve 0% survival under Byzantine conditions—their other-regarding mechanisms cause downward drift toward adversaries' zero contributions. Only unconditional commitment is structurally immune. v2.1.0 (2026-03-02): Rebuttal-hardened final version with 11 defense points. New: DNN ablation (4 architectures), K-Level Policy Gradients (K=0,1,2) comparison, 5×5 Spatial Social Dilemma, Nash Trap terminology precision (one-shot vs dynamic fixed point with p_surv derivative), principled design requirement framing. 13 pages, NeurIPS 2026 format. Code: https://github.com/openreview-neurlps/EthicaAI