Current multimodal forgery detection methods often struggle with generalizing across unseen manipulation types and lack transparent decision processes. To tackle these challenges, we introduces VeriChain–a novel reinforcement learning framework designed for explainable and generalizable document image forgery verification. Unlike traditional supervised fine-tuning approaches, VeriChain adopts a decoupled architecture that combines a Cognitive Reasoning Module with a Spatial Anomaly Segmentation Module. The Reasoning Module autonomously generates a self-revealing chain of forensic thoughts, highlighting multimodal inconsistencies and extracting spatially-grounded cues. These cues guide the Segmentation Module to precisely delineate pixel-level manipulation footprints. Trained exclusively via autoregressive policy optimization with a sophisticated dual-reward mechanism–enforcing structural coherence and forensic fidelity, VeriChain achieves emergent reasoning capabilities without explicit forensic reasoning data. This paradigm fosters robust zero-shot generalization across diverse and unseen manipulation techniques while delivering unprecedented interpretability through its self-derived reasoning chains. Comprehensive experiments on our proposed DocTruth benchmark, comprising over 126,000 diverse document images, demonstrate VeriChain’s state-of-the-art performance, achieving 90.8% detection accuracy, 78.3% localization IoU, and significantly superior explanation quality compared to existing methods. VeriChain establishes a new paradigm for trustworthy and adaptive document image integrity verification.