The national government has been increasingly vocal about the cadre fault-tolerant mechanism (CFTM), however, the spread of this policy in China has gradually slowed down in recent years. To solve the above problems, the key is to study the diffusion logic of CFTM. In the research on the diffusion logic of policy innovation, most research methods generally follow the "within-sample explanation" approach. This makes it difficult to ensure the scientific validity, especially the generalizability, of the research conclusions. To overcome this limitation, this paper adopts the "out-of-sample prediction" approach to explore the diffusion logic of CFTM. Specifically, based on the policy innovation diffusion theory, this paper constructs an analytical framework, uses machine learning methods to train a predictive model for local governments adopting CFTM, and presents the diffusion logic of CFTM while ensuring the predictive performance of the model. The model interpretation results reveal that the diffusion of CFTM is mainly dominated by actor logic, followed by efficiency logic and legitimacy logic. . Among the three dimensions of efficiency logic, legitimacy logic, and actor logic, the most influential features are governance scale, peer adoption, and the tenure of the top leader, respectively. The probability of local governments adopting CFTM is negatively related to governance scale, positively related to peer adoption, and has an inverted U-shaped relationship with the tenure of the top leader. To promote the orderly development of CFTM, this paper suggests that local governments strengthen the leadership-driven effect in the diffusion of CFTM and scientifically plan the diffusion path of CFTM.