B Theoretical Analysis B 理论分析The proposed world model is grounded in a unified understanding-generation-prediction substrate and a hybrid temporal backbone. To formally analyze its long-horizon consistency, this section investigates future targets requiring extended world-state information, such as object permanence, delayed physical effects, and multi-stage task variables. We address two central questions: when is a bounded recent window fundamentally insufficient, and under what conditions can a hybrid multi-scale memory recover near-Bayes-optimal prediction?所提出的世界模型建立在统一的理解-生成-预测基底和混合时间骨干之上。为了形式化分析其长时域一致性，本节研究了那些需要扩展世界状态信息的未来目标，例如物体永存性、延迟物理效应以及多阶段任务变量。我们聚焦两个核心问题：在何种情况下，有界的近期窗口在根本上是不充分的；以及在何种条件下，混合多尺度记忆能够恢复接近Bayes-optimal 的预测