Inherent quality control (IQC) of electronic medical records (EMRs) is a critical component in improving healthcare quality and patient safety. However, traditional manual quality control is inefficient and lacks standardized criteria. Although large language models (LLMs) have demonstrated remarkable potential in natural language processing tasks, their application to EMR quality control—where fine-grained structured output and complex logical reasoning are required—remains challenging. To address this issue, we propose M3-MedQC, a novel automated IQC framework. Specifically, we train the model based on the Qwen3 LLM using the Low-Rank Adaptation (LoRA) technique and design a multi-head classification architecture to handle multi-rule and multi-dimensional issue detection. More importantly, M3-MedQC incorporates a multi-granularity evaluation system, which not only assesses the model’s ability to identify problems at the triplet level, but also evaluates its overall compliance judgment at the record level. In the China Health Information Processing Conference (CHIP) 2025 shared task, our approach achieved second place with a final score of 70.78, providing an effective technical framework and evaluation system for future research on automated medical quality control.