With the rapid advancement of AI-generated video technology (Sora, Pika, MoRA), detecting synthetic videos has become a critical challenge. Current deepfake detection systems suffer from limited cross-modal analysis and lack of interpretability, despite their importance for forensic applications. This paper introduces a multimodal fusion framework that combines visual artifacts, audio-visual synchronization, and temporal consistency with explainable reasoning chains. Our approach achieves 82.1% test accuracy on the GenVidBench dataset (175,266 videos across 9 generation models) while providing verifiable chain-of-thought (CoT) explanations for each detection decision. We further integrate blockchain-based evidence logging to enable cryptographically secure audit trails for forensic investigations. Comparative analysis demonstrates multimodal superiority over single-modality baselines (visual-only: 78.1%, audio-only: 71.3%). The system achieves real-time inference (3.7s per 30-frame video on CPU) and maintains 76.3% cross-domain accuracy on unseen generation models—addressing key limitations in existing approaches.