Neural Bridge: Forecasting Future Human Arousal from Video-Derived Predicted Neural Response

Neural Bridge is a prospective affective-computing framework that forecasts where future human arousal movement and rare response events are likely to occur in unseen video, several seconds before the corresponding response occurs. The central result is a representation-to-task transformation. Neural Bridge operates on population-level predicted cortical response produced by the V-JEPA→TRIBE v2 video-to-neural pathway. Although that upstream system is highly capable at neural encoding, its raw predicted cortical output was weak for prospective arousal forecasting, achieving only 0.136579 PR-AUC on the fixed grouped AGAIN future-event target. Neural Bridge transformed that same substrate to 0.238341 PR-AUC, a 74.51% relative improvement on the same target and split family. The result shows that high-quality predicted neural response is not automatically a usable future-affect representation; Neural Bridge is the downstream system that makes it predictive. The mature system became substantially stronger again. Relative to the original validated continuous Neural Bridge, Phase 7 improved Spearman by 16.61%, top-5% future-movement lift by 23.59%, and top-1% lift by 14.52%. More importantly, the top-5% predictive margin added beyond an already optimized autoregressive persistence model increased by 98.92%, effectively doubling Neural Bridge's high-priority prospective contribution beyond response inertia. The full Phase 7 confirmation completed 420/420 prespecified evaluation cells, with positive results in 15/15 held-out fold-groups and 5/5 fold means. A separate prospectively locked cold-start study then tested the harder case in which no audience-response history exists at inference. On 299 previously unopened held-out videos, with no current arousal, past arousal, teacher score, or labelled warm start, Neural Bridge improved Spearman by 77.65%, top-5% lift by 70.80%, and future-event PR-AUC by 26.50% over the strongest endpoint-specific controls. All 5/5 panels were positive, with paired whole-video bootstrap lower bounds remaining above zero. Validation spans the complete controlling video corpora of 995/995 AGAIN sessions and 124/124 VEATIC clips, providing evidence across distinct stimulus domains and affect constructs. This record contains the journal-master preprint, complete LaTeX source, arXiv-ready source, research evidence package, and SHA-256 integrity manifest. The associated implementation is available through the linked GitHub repository.