Early and reliable short-video news verification through event-aware shot modeling and cross-modal evidence fusion

Short-video news combines titles, images, speech, and captions, allowing misleading narratives to emerge from individually authentic components. Whole-video aggregation and undifferentiated fusion can obscure when discrepancies arise and whether their sources are reliable. We present EASE-Fact, a framework for content-based news screening that models ordered visual units under an explicit observation boundary. A causal event-evolution transformer captures title-conditioned transitions, while a reliability-aware router models typed cross-modal relations and distinguishes inconsistent evidence from absent or unreliable sources. The Core model supports complete-video classification and a separately trained streaming configuration with prefix-local preprocessing and committed transcripts. The Full model adds temporally admissible historical debunking records and preserves the Core prediction when usable evidence is unavailable. Chronological and event-disjoint evaluations, matched comparisons, online first-crossing alerts, component ablations, and manual localization assess the framework’s performance and practical boundaries. The findings support temporal structure and source-quality modelling, while statistical uncertainty and event-exclusion analyses qualify comparative and retrieval gains. EASE-Fact provides inspectable diagnostics for news communication and editorial prioritization; benchmark predictions and internal conflicts remain distinct from independent determinations of factual truth.