TRACE-FND: A Constrained Evidence-Agent Workflow with Trainable Graph Adjudication for Multimodal Fake-News Detection

Multimodal misinformation often preserves individually plausible text and images while manipulating their relation, provenance, timing, or omitted context. Existing detectors mainly learn fused correlations or verify normalized claims, leaving temporal validity, source reliability, evidence conflict, and material omission weakly modeled. We introduce TRACE-FND, a constrained evidence-agent workflow with trainable graph adjudication for post-level multimodal fake news detection. Frozen structured agents convert each raw post into typed observations, truth/deception/omission hypotheses, and hypothesis-conditioned retrieval queries. Admissible candidates are normalized into immutable evidence cards; a trainable stance head and relation-aware graph discriminator then produce hypothesis-conditioned representations, while an evidence contract admits only card-cited arguments. A single calibrated judge maps the base, graph, hypothesis, debate, evidence-quality, and uncertainty features to the final label and confidence. The paper explicitly defines the frozen/trainable boundary, supervision, and one-to-one correspondence between mathematical objects and executable modules. Across Fakeddit-MM, Weibo-2017, and Multi-Fake-DetectiVE, the Teacher improves Macro-F1 over the strongest external baseline by 1.65, 1.48, and 1.60 points, respectively; paired tests support the first two gains, whereas the smaller Multi-Fake-DetectiVE test is not statistically conclusive. TRACE-FND also improves evidence quality, omission awareness, and temporal robustness, while a distilled Student reduces cached inference cost.