Detecting texts generated by Large Language Models (LLMs) remains highly dependent on the availability of comprehensive training datasets. In response to NLPCC 2025 Shared Task 1, this paper introduces a low-cost yet effective detection system that combines traditional machine learning techniques with GECScore-based grammar correction. Our method demonstrates strong performance across both unseen domains and attacked data, all without the need for data augmentation, achieving an average F1 score of 94.94%. Key features used in our approach include entropy, perplexity, Chinese-English code-switching, special character usage, clause length, repetition, sentiment polarity, and grammar correction. Among the models evaluated, Random Forest emerged as the top performer. The proposed solution is both cost-efficient and resilient, offering reliable detection of LLM-generated content.