Planner Objectives Rather Than Predictor Accuracy Limit Latent World Models

In latent world model evaluations, failures in long-horizon planning are caused by suboptimal planning objectives rather than predictor degradation. On the TwoRoom environment, ridge probes recovered position from frozen embeddings with R^2 0.9922, proving the necessary information exists within the latent state but is unrecoverable by current cross-entropy-method planners.