The phone is propped so it cannot move. Five or more objects go on a table; one is removed and named between each of five frames. Sequences where the camera moved, two objects left at once or a removal went unnamed are rejected, one sequence per contributor. Each sequence is a four-step chain with exactly one named object leaving per step, so the decrement, not the absolute count, is ground truth, on real tables under real light, with shadows and reflections synthetic edits rarely get right. Counterfactual image pairs, identical except for one known change, are what change detection and visual reasoning evaluations want, and are rare as licensable data.