<p>This study addresses program failure diagnosis by analyzing fault propagation paths and predicting fault behaviors through clustered “Golden Runs” (error-free executions). Key challenges include accurately characterizing fault-triggered program behavior and efficiently tracing the “fault-error-failure” chain by reducing complex fault behavior spaces. Guided by the hypothesis that workloads with similar attributes exhibit comparable fault behaviors, we propose a method to profile and cluster “Golden Run” patterns to predict fault manifestations. Central to our approach is the Fine-Grained Basic Block (FGBB), a runtime state metric combining instruction addresses and fault manifestations to detect initial anomalies during fault activation. We validate the method through extensive fault injection experiments on SPEC CPU2006 benchmarks. Results demonstrate that clustered “Golden Runs” achieve up to 73% similarity in fault behaviors within the same cluster, particularly when fault/failure categorization is optimized. The FGBB framework identifies 75% of faults before full propagation, improving diagnostic efficiency by 40% over conventional methods. By anchoring fault prediction in normal execution patterns, this work reduces computational overhead while balancing precision and efficiency. The clustering-driven mechanism compresses fault behavior spaces into manageable models, enabling scalable failure analysis in complex systems. This approach provides a systematic methodology for root-cause diagnosis and reliability assurance in resource-constrained environments.</p>