A video benchmark for object permanence. Each sample pairs an input clip with its ground-truth continuation while objects move, vanish, or persist behind occluders. Dataset v1.2.0.