Researchers have introduced Temporal-Distance-JEPA, a novel representation learning framework designed to enhance latent Model Predictive Control (MPC) within Joint-Embedding Predictive Architectures (JEPAs). Traditional JEPA planners often rely on latent Euclidean distance for ranking imagined futures, which may not align with actual task progress. To address this, the authors propose mining a directed temporal cost directly from reward-free offline demonstration logs. This method utilizes three specific supervision signals: same-trajectory step order provides positive targets, cross-trajectory pairs serve as heuristic negatives, and a rollout-consistency term is employed to match the specific planner horizon. The resulting representation serves a dual purpose: it functions as a topological planning cost and improves Euclidean planning in environments dominated by contact geometry. Experimental results demonstrate significant performance gains over the LeWM baseline. Specifically, deploying the mined cost achieved a 100.0% success rate in the Two-Room environment compared to LeWM's 97.4%. Furthermore, using shared Euclidean planning on the temporally trained checkpoint increased the OGB-Cube score by 14.2 points over LeWM and improved performance in Push-T tasks. Ablation studies confirm that the directed head, cross-trajectory negatives, and rollout consistency are all essential components for narrowing the gap between representation training and effective deployment in JEPA-based world model planners.