Calibrated Typed Relations from World Model Latents for Relational Composition

World models organize predicted futures in latent representations, but prediction alone does not expose the relations needed for downstream reasoning. This paper studies an explicit relational layer over frozen world model representations. A fixed graph separates structural admissibility from contextual activation, while source resolved evidence places uncertainty on typed operational relations rather than on the latent state as a whole. Compact relation records carry calibrated activation, predicted response magnitude, and horizon. Local relational composition then tests compatibility at a shared node. A paired complex encoding assigns relation type and edge orientation to fixed complex axes, while a matched real encoding preserves the same relation content. Across two frozen MetaDrive evaluation splits, structured evidence supports strong relation inference and simple aggregation remains competitive with learned aggregation. The paired complex encoding shows a modest composition gain, while ensembling improves composition discrimination and future latent prediction. The results support separating relational uncertainty from compositional compatibility and the frozen predictive backbone.