Can Predicted Dynamics Exist in the Physical World?

This paper asks whether a predicted rollout, action chunk, or latent plan can be physically executable before execution, and argues that low prediction error is not enough. The contribution is a model-agnostic physical-admissibility gate that decomposes executability into flow consistency, recursive reachability, bounded differential growth, and learned dynamics consistency, then rejects any decoded proposal whose worst normalized residual exceeds a threshold. The gate is evaluated on Hugging Face LeRobot PushT, using compact MLP world-model baselines with a horizon of 32 and controlled falsification families including smooth impulse, actuator lag, time warp, mode change, action-state mismatch, and action saturation. On dynamic-violation detection, the strongest scalar residual is the transition-RMSE detector with AUC 0.982 and AP 0.997, followed by the standardized dynamics residual with AUC 0.972 and AP 0.995; kinematic-only scoring is much weaker at AUC 0.592, while the full admissibility gate reaches AUC 0.957 and AP 0.993 with condition-level attribution. In replay intervention, the full physical gate prevents 87.7% of invalid proposals while causing 8.5% false interventions, and retained nominal progress stays near 0.998. The paper also shows that history-conditioned prediction is more accurate than a state-only Markov ensemble on PushT, with rollout RMSE 0.00221 versus 0.01000, exposing partial observability in the monitored state and motivating runtime verification at the prediction-control interface.