Dismantling the Soft-Law Illusion and Building Minimum Substantive Control Capability (MSCC) for Public-Sector AI Governance

This study addresses the profound governance crisis arising from the deployment of Generative AI and Retrieval-Augmented Generation (RAG) technologies within the public sector. For years, Taiwan's reliance on "soft law governance"—spearheaded by soft-guidelines and self-compliance checklists drafted by institutes like the Science & Technology Law Institute (STLI)—has fallen into irreversible "triple systemic shackles": the proprietary boundaries of commissioned institutes, heavy reliance on IT vendors, and the creation of a bureaucratic comfort zone for de-responsibility. Technologically, generative AI induces a "high structural decoupling" of administrative power, characterized by outsourced GPU infrastructure, probabilistic hallucination, black-box RAG vector slicing, and a "responsibility evaporation chain" driven by automation bias.

To dismantle this constitutional and administrative predicament, this research proposes the "Minimum Substantive Control Capability" (MSCC) as a hard-law verification standard for public sector AI deployment. MSCC is operationalized into four pillars: Zero Data Retention (ZDR) lineage, unilateral Kill-Switch mechanisms, chunk-level attribution, and immutable full-lifecycle log auditing. Furthermore, this study constructs a tri-agency implementation matrix: MODA standardized procurement contract clauses, Audit Court full-lifecycle digital audit checklists, and Judicial review standards featuring reversed burden of proof. This architecture fills a crucial academic and practical void in public administration and administrative law, establishing operable, quantifiable, and enforceable supervision over black-box algorithmic governance.