Friction-MARL: 5×5×5 Factorial Results (Multi-Agent Coordination under Stakes, Alignment, and Observation Noise)

Complete results of a 5×5×5 factorial multi-agent reinforcement learning (MARL) study of coordination among four Independent Q-Learning (IQL) agents managing three shared, capped, continuous resources. Reward is the negative weighted squared distance from each agent's ideal resource state. The factorial varies three parameters, each over five levels, with 30 independent replications × 1,000 episodes per condition (125 conditions): alpha (−0.8, −0.4, 0.0, 0.4, 0.8) — preference alignment: the correlation structure of agents' target preferences (cooperative / unrelated / adversarial). This is the structure of disagreement, not a friction magnitude. sigma (0.2–1.0) — preference intensity / stakes: how much agents care about each resource. epsilon (0.0–1.0) — observation noise: Gaussian noise on each agent's observation of the resource state. Two independent implementations (CPU multi-worker and GPU-vectorized) were run and cross-validated (Spearman ρ = 0.937 rank-order agreement). Headline findings: a U-shape in alpha (structured disagreement — cooperative or adversarial — beats unrelated preferences); stakes dominate alignment structure; friction equalizes cross-agent outcomes; reward convergence occurs without policy convergence; and observation noise has near-zero effect. Contents: per-condition and per-replication CSVs, per-replication learning curves and final policy vectors (.npz), the full statistical analysis suite (heatmaps, residual/model-comparison plots, ANOVA and regression tables, dynamics/clustering figures), the GPU–CPU cross-validation outputs, paper-ready LaTeX tables, analysis reports, and the runner/analysis code (friction_marl package). See README.md for the full directory layout and column schemas. This dataset is the data-availability source for the MARL companion paper (DAI-2606, When Stakes Dominate, forthcoming) and provides the empirical grounding for the friction operator in the Axiom of Consent (arXiv:2601.06692) and the Replicator Optimization Mechanism (arXiv:2601.06363).