Public administrations traditionally operate on deterministic, rules-based systems, which increasingly face scalability challenges when dealing with complex, unstructured governance tasks. This paper evaluates the transformative potential of Large Reasoning Models (LRMs) and the architectural paradigm of "Nested Learning" within Bavarian public governance. While standard Large Language Models (LLMs) often lack the precise legal logic required by statutory obligations such as the administrative duty to provide reasons (§ 39 BayVwVfG), emerging "System 2" AI models offer human-like, multi-step inference pathways. This study develops a sovereign framework that integrates Robotic Process Automation (RPA) with LRMs to serve as a reliable "subsumption engine" for civil servants. Furthermore, it addresses critical socio-technical vectors, including mitigation strategies for automation bias, fiscal sustainability, and data sovereignty under the EU AI Act and GDPR frameworks. The findings indicate that leveraging inference-time scaling models can effectively bridge the gap between static bureaucracy and agile, AI-ready public administration without compromising legal certainty.
Aaron Rinberger (Mon,) studied this question.