Decision-Level Auditing of Latent Robot World Models

Before adapting a robot world model, it is useful to test how terminal-prediction errors affect action selection under its existing objective. We introduce a reproducible decision-level audit that replaces predicted terminal latents with simulator-realized endpoint encodings while fixing the candidate actions, encoder, goal, and scoring objective. Using released JEPA-WM and DINO-WM checkpoints in Push-T and Wall, we measure changes in selected actions and their continuous physical outcomes, with task completion retained as a separate outcome. Predicted scoring exceeded a uniform menu baseline on at least one physical outcome in each environment. In Wall, endpoint substitution changed 100/179 valid choices, yet its native-utility effect across 29 complete cases was $-.254$ (95% CI $[-.909,.364]$; Holm-adjusted $p=.923$). This does not establish an additional improvement or equivalence. A Push-T-frozen near-tie diagnostic transferred prospectively to Wall, with AUROC $.727$ versus $.735$ for a raw-margin baseline. A CPU recomputation artifact makes 53,700 model-candidate records inspectable and reproduces the primary Wall analysis. The audit informs evaluation before adaptation; it does not simulate a learned post-training update or isolate representation and objective effects. Findings are limited to fixed initial menus, open-loop action sequences, and predictors sharing DINOv2 features.