PAN 2 opened a world model's weights and code. Keep the physics test open

THE SIGNAL IN ONE SENTENCE

A video can look exactly like physics right up to the moment the pasta appears in the colander before anybody drains the pot. That example comes from the limitations section for PAN 2, a newly released world model from the Institute of Foundation Models, the research institute established by Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi. On November 10, the institute released PAN 2 as an open system for simulating what could happen after a described action. Give it text, and optionally a starting image, and the system produces a video up to 20 seconds long. The idea is different from ordinary prompt-to-video entertainment. A world model is supposed to represent how a scene changes. An agent could imagine several candidate actions, inspect the simulated consequences and then choose what to do in the real world. PAN 2 splits that job in two. A language model handles conceptual reasoning. It writes a structured, time-grounded account of events, actions, actors and camera movement. A 29-billion-parameter latent diffusion model handles perceptual reasoning. It turns that plan into moving pixels at 704 by 1,248 resolution and 24 frames per second. The release is unusually inspectable. The institute published the video-model weights, training code, inference code, documentation and technical material under the Apache 2.0 license. The repository supports text-to-video and image-to-video generation, and its model card describes the separate third-party components and licenses used in the pipeline. The institute reports that PAN 2 completed 75.1 percent of individual steps and 34.0 percent of full plans in its multi-step evaluation. It says that was the highest full-plan completion among the models it studied. It also reports third place overall and first among open-weight entries on the public Physics-IQ Verified leaderboard setting. Those are the releasing institute's results, not an independent replication. They are also a useful reality check. A 34 percent full-plan score means about two out of every three evaluated plans contained at least one missed step. That is promising research and a very bad autopilot. The open release gives other researchers the ingredients for a better question: can the simulation survive an open physics test? Start by preserving the complete test case. Save the starting image, original request, rewritten structured prompt, negative prompt, conceptual reasoner, model checkpoints, software versions, random seed, sampler settings, output resolution, clip duration and hardware. Keep the generated video and the configuration file beside it. That matters because PAN 2 is not one sealed box. The conceptual reasoner can change. The institute's own comparisons used different language models in different settings. The public Physics-IQ entry used Claude Opus 5.5 as the conceptual backbone. Another reported benchmark used GPT-6 Astra. The technical report says its main studies used Kimi K3 unless otherwise stated. If the language model silently rewrites the experiment, a score can reflect the planner, the video model or both. Report them separately. Next, write the expected constraints before looking at the output. If a ball falls from a shelf, gravity should move it downward before contact. If oil enters water, it should not behave like ink. A closed drawer should not let an object pass through its front. A cup should not refill between camera angles. A hand should not gain fingers to finish a grip. A robot should not reach through a barrier because the final frame looks tidy. Turn each requirement into a timestamped check. Did the event happen? Did it happen in the right order? Did the relevant objects persist? Were mass, contact, containment, support, momentum and material behavior plausible? Did a camera move reveal the same room, or invent a convenient replacement? Do not score only the last frame. A system can land on a plausible result by taking an impossible route. PAN 2's own model card warns about physical shortcuts in intermediate states. Those shortcuts are exactly what a robotics or planning system cannot afford to ignore. Use real controls. Record a real version of the same tabletop experiment where practical. Run more than one real take because physical outcomes vary. Compare the generated trajectory with the distribution of observed outcomes, not with one lucky clip. For tasks that cannot be staged, use established simulators or measured datasets, but record where the reference itself simplifies the world. A simulator that ignores friction is not a magical source of friction truth. Then vary one thing at a time. Change the material, starting angle, mass, surface, action wording, camera position or lighting. Keep everything else fixed. A useful world model should respond to the physical change without collapsing under an irrelevant rephrasing. PAN 2's model card says prompt wording matters and that some demonstrations outside its training data used hand-refined prompts. That does not make the demonstrations useless. It means the prompt-editing history belongs in the evidence packet. Run several seeds. Report the median, spread and worst meaningful failure, not only the best-looking video. If one seed obeys the constraint and four spill through the table, the open test should not publish the photogenic winner as the result. Label failures in a shared vocabulary. Useful categories include missing action, wrong order, entity substitution, identity drift, spontaneous object, disappearing object, broken contact, impossible penetration, incorrect material behavior, camera-induced world change, unsafe action, and plausible ending reached through an impossible intermediate state. Keep the raw labels and reviewer notes. A single physics score is handy for ranking. It is lousy at telling a builder whether the model failed because a ball floated, a person changed identity or the camera hid the missing step. Human review needs a protocol too. The institute reports a blinded comparison using 140 prompts, 13 reviewers and 1,400 pairwise judgments across action adherence, physical plausibility and world consistency. That is more informative than a highlight reel, but human viewers can still reward visual polish. For each test, ask reviewers to score the physical constraints before overall preference. Include slow playback and frame stepping. Let reviewers mark uncertainty. Measure agreement. When they disagree, keep the disagreement visible instead of averaging it into calm authority. Separate appearance from prediction. A world model can be useful for ideation, education or synthetic scene generation even when it is not accurate enough for physical control. The release's own intended uses include research, robotics, education and creative work. Its out-of-scope guidance says it should not make safety-critical decisions without independent human review and validation. That boundary should travel with every demo. A simulated clip is not evidence that an event happened. It is not a forecast that the same action will produce the same result. It is not permission for a robot, vehicle, medical device or industrial controller to act. Open weights make inspection possible. They do not turn generated pixels into measured physics. Safety needs an open test as well. The PAN 2 model card recommends a separately distributed guardrail that checks requests before generation and generated frames afterward. In the institute's internal evaluation with Kimi K3 as the conceptual reasoner, the guardrail reduced harmful text-to-video requests that produced unsafe video from 33.9 percent to 1.5 percent. It also blocked 11.5 percent of benign text-to-video requests and 3.3 percent of benign image-to-video requests. Those figures show both sides of a filter. Harmful output fell sharply in the reported test, and legitimate requests were sometimes stopped. Independent evaluators should rerun the test across languages, cultures, adversarial prompts, image inputs and software versions. Publish false negatives and false positives together. The release also creates a provenance problem familiar to every realistic video generator. A convincing simulation can be mistaken for footage. The model card warns against misinformation, impersonation and non-consensual or harmful content, and asks users to disclose generated material where it could mislead. Projects built on PAN 2 should add durable provenance metadata, visible disclosure, content handling rules and a route for reporting misuse. Open code makes adaptation easier. It also makes a safety wrapper easier to remove. Hardware access matters. The public checkpoint is 86.2 gigabytes on Hugging Face, and the perceptual backbone runs in BF16. The release documents a serious research stack, not a toy model for an ordinary laptop. Reproduction reports should include accelerator type, memory, runtime, energy where measurable and the cost per generated test case. That information changes who can actually participate in open evaluation. A license may be open while meaningful testing remains expensive. The strongest part of PAN 2 is not a claim that physical intelligence is solved. It is the opportunity for other researchers to open the box, reproduce the pipeline, change the test and publish where the simulated world breaks. The institute itself says long-term consistency remains an open challenge. The model sees at most one starting frame plus text for each clip, so appearance can drift across longer interactions. Text is also a coarse way to specify dexterous motion, such as exact finger contact. The current quantitative study focuses on open-loop plans designed to succeed and does not yet show how faithfully the model handles plans that should fail. Those are not footnotes to hide under a cinematic demo. They are the research agenda. Keep the physics test open. Publish the starting state, expected constraints, prompt history, reasoner, seed, hardware, output, real-world control, failure labels, reviewer disagreement and guardrail result. Then let somebody else run it. The interesting world model is not the one that makes the most persuasive video. It is the one whose mistakes become measurable before anybody asks it to touch the world.