Inputs are 224 × 224 RGB observations, a goal image, and the causal history of the previous five primitive actions. The ViT-Tiny encoder starts without pretrained weights and uses a 192-dimensional latent space. The evaluation uses three training seeds (3072–3074) and three evaluation seeds (0, 1, 42), with 100 episodes per evaluation seed and task on an RTX 4090. Goals are 25 steps ahead, with at most 50 actions, a planning horizon of 5 and replanning every 5 actions. Four-task macro success is 94.8 ± 0.4% (Direct) and 95.1 ± 0.3% (Guarded A). Separately trained models score 94.9 ± 0.5% and 95.6 ± 0.9%, respectively. The values are means ± standard deviations across training seeds. Inference uses FP32 with solver batch size 1. Direct generates the actor’s plan without candidate search. Guarded A evaluates 128 samples per iteration for three iterations, using the top 16 and an initial standard deviation of 0.25. These numbers describe the released Shared BatchNorm checkpoints; the paper’s other ablation protocols report separate results.