Detecting financial misinformation without external references is a critical challenge, as large language models (LLMs) often exhibit a compliance bias that causes them to blindly accept mathematically coherent but contextually false claims. This paper describes the submission by Team Coherence for the Financial Misinformation Detection (FMD) 2026 shared task. To overcome LLM compliance bias, we leverage a courtroom-inspired framework. For a given financial claim, our system utilizes the DeepSeek-chat model to generate adversarial prosecution and defense arguments. These generated perspectives, alongside the original claim, are then evaluated by a multi-channel classification network featuring a frozen ModernBERT backbone, parallel BiGRU layers, and context-based attention pooling to optimally weigh the evidence. Crucially, our technique does not rely on the specific perturbation details of the dataset, allowing it to maintain generalizability. Evaluated on the blind test dataset, our framework achieved an accuracy of 0.955 and an F1 score of 0.9549, securing third place on the FMD Challenge 2026 leaderboard and demonstrating its viability as an efficient, objective judge for reference-free financial misinformation identification.

PDF