Operational Change as a Representational Primitive for World Model Latents

World models encode histories of observation and action in latent states that support prediction and control, but predictive utility does not determine how changes relevant to decisions should be represented. This paper asks whether a frozen world model latent can directly support independent representations of increase and decrease for five broad decision quantities, namely risk, value, opportunity, certainty, and belief. Rather than decoding one scalar for each quantity and deriving change by subtraction, the learned projector predicts support for directions of operational change from one current latent. Increase and decrease are not constrained to occupy opposite ends of one scalar axis, and both can receive support when local interventions produce opposing changes. Continuous targets are constructed from exact local interventions, complete simulator states, and oracle posterior updates while privileged information remains outside the learned projector. Across five seeds under a fixed test protocol, Independent Directional Nodes reduce macro node mean absolute error by 41.5% and pair total mean absolute error by 50.0% relative to Signed Axes of matched capacity, while increasing order concordance by 0.203. Relational fidelity further increases order concordance by 0.046 relative to Independent Directional Nodes, with only a small pointwise fidelity tradeoff. These results support representing directions of operational change directly rather than compressing opposing directions into a single signed coordinate.