A Multi-Agent AI-Blockchain Framework with Reverse Kelly AMM for Under-Collateralized Real-World Asset Lending

We introduce an Automated Market Maker (AMM)-based lending mechanism that applies a Reverse Kelly criterion to establish loan premiums based on model-estimated default probabilities and collateral ratios. We embed this mechanism in a multi-agent system that completes the financial loop using on-chain reputation and enforcement. This specific combination of Kelly-optimal credit pricing with multi-agent orchestration for under-collateralized assets constitutes the core novelty of our research. Unlike traditional AMMs designed for token exchange, we adapt Kelly’s growth-optimal allocation principle to the credit market, thereby establishing a dynamic pricing surface that explicitly links premiums to probabilistic risk. The AI-Blockchain Reverse Kelly AMM (rkAMM) framework integrates three types of autonomous agents: (i) AI-based Risk Assessment Agents (RAA) that estimate borrower default probability (PD); (ii) AMM Pricing Agents (PA) that use the Reverse Kelly criterion to determine loan premiums and an optimal capital allocation fraction; and (iii) Smart Contract Enforcement Agents (SEA) that guarantee transparent execution and update an immutable on-chain reputation registry. We provide empirical validation of the framework through extensive simulation. First, a head-to-head ablation study on an identical 10,000-loan stream demonstrates that the Reverse Kelly strategy achieves a 14.3% annualized growth rate, which surpasses proportional-premium (10.8%) and fixed-premium (7.6%) models. We report critical risk metrics, including max drawdown (18.2% vs. 25.4%) and loss ratio (8.1% vs. 12.7%), confirming superior risk-adjusted returns. Second, a stress test utilizing fat-tailed PD shocks verifies that the Kelly-based allocation rule automatically clips exposure, which maintains pool stability. Finally, a minimal Layer 2 (L2) testnet deployment validates the proposed on-chain logic. Our results offer robust and reproducible evidence for this novel, capital-efficient, and resilient architecture intended for decentralized under-collateralized lending.