Why can some local reforms cross the “minefield of implementation” while others are fragmented among departments and levels? This paper takes the healthcare reform in Sanming， China， which has attracted much attention due to the continuous cross-level collaboration， as an example.It uses a generative AI approach to delve into the “black box” of policy implementation， and establishes a framework for the consistency evaluation of public policies from four core elements： idealized policy， policy implementers， target groups and policy environment， revealing the dynamic mechanism of the consistency evaluation of public policies.The research findings are as follows： (1) The distribution and evolutionary trends of thematic terms concerning medical care， medical insurance and pharmaceuticals in the Sanming healthcare reform shows strong correlations， indicating a high level of public policy coherence.(2) Through DTM analysis， the Sanming healthcare reform has undergone a three-stage dynamic evolution path from idealized policy to policy implementers， then to target groups combined with policy environment.(3) By applying large language models to mine the four dimensions of the public policy consistency evaluation framework， a systematic element deconstruction table was formed， providing a practical basis for promoting the consistency evaluation of public policies.