Governance-aware autonomous retail coordination in artificial intelligence cities using multi-agent reinforcement...

When autonomous systems take operational authority over urban commerce, accountability and human oversight matter as much as efficiency. We present AAIRM, a governance-aware autonomous retail coordination framework addressing three requirements for trustworthy procurement: tamper-evident decision provenance, data sovereignty, and adaptive demand coordination. AAIRM couples a Proximal Policy Optimization (PPO) ordering policy, a permissioned Blockchain Trust Ledger (BTL) for multi-party audit trails, and a Federated Demand Learning (FDL) layer. Evaluation is simulation-based, using a multi-category synthetic environment and the public M5 dataset; no live retail deployment is claimed. At its cost-optimal operating point (92.3% fill), AAIRM lowers normalized inventory cost by 13.2% (95% CI 12.2-14.2) against a conventionally parameterized reorder-point/economic-order-quantity (ROP-EOQ) baseline and by 10.2% (95% CI 8.0-12.4) on M5. Service parity is obtained by moving AAIRM along its own cost-service frontier rather than by retuning baselines: At 96.1% and 97.4% fill, it retains 8.2-point and 4.9-point advantages. A layered ablation attributes 8.8 points of the advantage to the reinforcement learning (RL) policy and 3.8 points to the combined coordination-and-governance block, which we further decompose into feasibility projection, escalation logic, cross-category policy rules, supplier ranking, and negotiation. The language-model orchestrator contributes 0.6 points (95% CI -0.9 to 2.1), statistically indistinguishable from zero and practically equivalent within a ±2-point margin. Against a feasibility-matched multi-agent RL comparator, the residual cost gap does not survive multiple-comparison correction, so the defensible advantages are hard feasibility and auditability, not cost. The BTL adds 6.6% decision-cycle overhead; its 500-case mutation replay is an integration check, and we give an explicit adversary model showing that crash-fault-tolerant ordering does not resist colluding majorities. FedProx converges within 0.9 WAPE points of centralized training at a 1.3-point cost penalty, without secure aggregation or differential privacy; we analyze rather than evaluate when those mechanisms become necessary. A 500-SKU study exposes a reward-hacking failure in a high-spoilage category, corrected by category-specific governance. We close with an architectural, non-empirical reading of labor exposure and three governance instruments regulators can act on.