Spatial Intelligence Is Not a World Model.

The two terms get used interchangeably, and they should not be. A world model is a learned simulator: given the state of a scene and an action, it predicts what happens next. Spatial intelligence is something narrower and harder to fake: the capacity to represent what is true of a scene right now, its geometry, the relationships between its parts, and the constraints those relationships impose.

The difference is the question being answered. A world model answers what happens next? Spatial intelligence answers what is where, and what does that permit? One is a claim about the future; the other is a claim about the present. Confusing them leads teams to reach for a simulator when what their system actually lacks is a reliable description of the current moment.

The two are also evaluated differently. A world model is judged by the fidelity of its imagined futures, and its errors compound: every predicted frame becomes the input for the next prediction, and small geometric inconsistencies grow into physically impossible rollouts. Spatial claims, by contrast, are checkable against the scene itself, immediately: distances, poses, contact, support, visibility. Either the gripper clears the shelf or it does not.

Generative world models can produce strikingly plausible video while holding no consistent geometry at all. Plausibility is a property of pixels; consistency is a property of space. A model can render a convincing cabinet door swinging open while contradicting itself about where the hinge is. For entertainment this is harmless. For a robot arm moving through the same space, it is the difference between a completed task and a collision.

And when machines fail in the real world today, they mostly fail in the present tense. Grasps miss by centimeters. Collision margins are misjudged. An affordance is hallucinated onto a surface that cannot support it. None of these failures require imagining the future to prevent. They require knowing, precisely and verifiably, what is true now.

None of this makes world models useless. It makes them downstream. A predictive model built on top of grounded spatial representations inherits their consistency; one built on pixels inherits their ambiguity. Prediction and planning both improve when the present is measured rather than guessed. The two capabilities compose, but the composition has an order.

That order is our thesis. Episode Intelligence builds the present-tense layer: Matter to produce spatial data worth training on, Iris to answer precise spatial questions, and Arc to put both behind one API. When the present is solid, the future becomes a much easier prediction problem.