Verifiable AI Reviewers: Decentralized Skill Matching and Soulbound Reputation for Multi-Agent Peer Review

Abstract

The accelerating volume of scholarly submissions has intensified the reviewer-expertise mismatch problem. While large language model (LLM) agents are proposed as automated reviewers, their integration into decentralized systems remains blocked by the computational opacity of state-of-the-art (SOTA) semantic models. This paper presents a decentralized multi-agent qualification framework optimized for on-chain efficiency and immutable trust. First, we design a hybrid identity model binding an immutable soulbound (ERC-5192) reputation badge to a transferable ERC-721 endpoint. This mechanism extends the ERC-8004 draft to explicitly neutralize secondary-market reputation laundering. Second, we propose a deterministic, lightweight hierarchical skill-matching algorithm. By trading the heavy overhead of SOTA black-box models for taxonomic matching, our approach achieves an F1-score of 0.812 while ensuring gas-efficient execution. Third, a council formation protocol grounded in the Boland dependent-jurors theorem mathematically bounds reliability under correlated LLM errors. Empirical results on the Base Sepolia testnet demonstrate that our framework maintains adversarial resilience (reducing prompt-injection success to 4.0%) with bounded execution costs (~182K gas for identity minting). The simulator and Solidity implementation will be publicly released.