PAI-ECON-CLAUDE: A GATED HUMAN-IN-THE-LOOP MULTI-AGENT ARCHITECTURE FOR AI-ASSISTED ECONOMIC THEORY DEVELOPMENT
创建于 更新于
摘要
This paper proposes pAI-Econ-claude, a gated, human-in-the-loop multi-agent architecture for AI-assisted economic theory development, integrating inspectable intermediate records and targeted quality gates. [page::0] The design emphasizes interception at visible decision points, non-certifying gates, and irreversible human checkpoints to improve auditability without replacing formal verification. [page::0] In a matched-task evaluation across five economic-theory challenges, the gated workflow outperformed ungated baselines on four tasks, with substantial reductions in reported failure severity and higher usefulness. [page::0] A notable success occurred when a reality check rejected a false market-structure premise and a subsequent proof review corrected a welfare claim, while Task 1 shows limits where extra structure may compress key mechanisms. [page::0] The results support the claim that bounded gated oversight enhances interpretability and auditability of AI-assisted theory, with human judgment concentrated where irreversibility is highest. [page::0] The code and workflow are publicly available at the linked GitHub repository. [page::0]
