Multimodal Futures and Calibrated Uncertainty in JEPA Predictors

A deterministic predictor in embedding space collapses multiple valid futures into their average, which can be a state that never occurs. We study whether a JEPA predictor can represent several futures, and whether its uncertainty is calibrated. We use simulated systems with known branching, and a language setting in which a prompt admits several correct continuations. A bimodal pilot shows the mixture interval attains its nominal coverage; the JEPA results are pending. Preprint. Results in this version come from small-scale pilot experiments; the full experimental protocol is described in the text and is pending.