High-quality electronic medical records (EMRs) are fundamental for patient safety, clinical decision-making, and precision medicine. However, traditional manual quality control is inefficient, inconsistent, and inadequate for the growing complexity of clinical data. We introduce a large language model (LLM)-based framework that combines parameter-efficient Q-LoRA fine-tuning with task decomposition to enable automated, rule-specific validation of EMR content. This approach leverages iteratively optimized instruction prompts to guide the model through focused single-rule checks, thereby enhancing both accuracy and explainability. Evaluated on a curated dataset using the Qwen3-32B model, our method achieved a comprehensive score of 0.6918, demonstrating a substantial performance improvement over untuned baselines. By integrating intelligent instruction generation, modular task decomposition, and scalable fine-tuning, this framework provides a robust and adaptive solution for clinical data quality assurance. Our study underscores the potential of combining large-scale generative models with domain-specific adaptation to transform medical data governance and set new benchmarks for automated, high-fidelity clinical data evaluation.