Timing decides outcomes: a surfer has to pop up just as the wave lifts the board, and too early or too late, the ride is lost. PAN 2's conceptual reasoner writes out who does what and when, so actions can take effect in the order given. With PAN 2, you decide when each action happens. Change the timing or the order, and you see a different outcome, with each step building on the one before. Only a simulation faithful across the full plan can reliably guide an agent's decisions.

Pour milk, ink and oil into water: milk clouds it, ink spreads through it, and oil floats on top. An agent needs a simulation that obeys such physical laws and constraints, not just a visually convincing clip. PAN 2's perceptual backbone fills in what the conceptual backbone leaves open: geometry, appearance and the details of motion. Trained against real footage, it learns how the physical world evolves. It answers “what if?”, even for combinations of conditions absent from the training data.

Turn the camera full circle, and the room you return to is the one you left. An agent's view changes with every move it makes, but the world it acts in should not. PAN 2's conceptual reasoning backbone writes camera motion into each next state, separately from what people and objects do, so the camera can move without changing the world it shows. What leaves the view stays in the world, keeping its identity and state until it comes back.