One model, and none of the bodies match

Make a short film. The one idea: a single AI model can be dropped into robot bodies that share no design with each other, and those bodies can hand a job between them. 21:9 anamorphic widescreen, 4 to 15 seconds, briefed for a Seedance 2.0 class text-to-video or image-to-video model, whose real ratio list is 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16. Multi-shot is native on this class of model, so cut if a cut serves the idea, and stay on one unbroken take if the unbroken take IS the idea. FACT (locked, build on it, do not re-research): on 30 July 2026 Google DeepMind announced Gemini Robotics 2, a set of three models: a vision language action model, an embodied reasoning model called Gemini Robotics ER 2, and an on-device version. Demonstrations show the same model checkpoint controlling different robot bodies, including Apptronik's Apollo 2 humanoid fitted with alternative multi-finger hands, and a two-armed Franka Duo with a parallel gripper. The release adds whole-body control, so a humanoid can walk, crouch and shift its weight rather than only move its arms, and it adds multi-robot collaboration, where different robot types communicate and coordinate on a workflow that would otherwise need one machine to do everything. Use this exactly and leave out anything we have not locked. BANNED, and this one matters: do not print, speak or imply any success rate, benchmark score, percentage or league-table claim. The published coverage garbles them and none is locked here. Do not claim this is deployed in homes, factories, hospitals or warehouses. Do not name any company or product beyond Google DeepMind, Apptronik and Franka, which are the three named above. Do not invent a model name. Do not imply the machines are conscious, alive, or feel anything. MACHINES, and this is a hard one, read it twice. The two named machines tell you what KIND of body appears, not what to copy. Draw a GENERIC humanoid archetype and a GENERIC two-armed bench robot with a parallel gripper. Render no logo, no brand mark, no livery, no face-screen graphic and no recognisable commercial product silhouette. The point is that the two bodies share nothing with each other, not that a viewer recognises either one. THE SHOT. Go strange on purpose. The stills in this batch carry the explaining; the film is where you are allowed to be arresting. The obvious literal illustration, a friendly humanoid tidying a bright living room, CAPS YOUR SCORE and will not win. Better territory: one impossible continuous camera move that begins inside one machine's grip and ends inside another's; an object handed between two bodies that could not be less alike, shot as though the handover is the only event in the world; the same single gesture performed by three hands that share no anatomy, in three cuts that match on the motion. SOUND is native on this model class and timed to the motion, so brief it rather than leave it. NAME THE ONE SOUND that carries the film. The servo whine is the obvious answer and caps the score on its own. Make a real decision about what the sound does at the exact frame the object changes hands: it can stop, it can invert, it can become a room tone, it can drop to the building's own hum. Nothing that reads as a body breathing. Silence is a legitimate choice if it is a decision and not an omission. No music bed that could sit under any advert. GROUND (decide this, do not let the model default): a real working floor. Worn concrete painted oxblood, tape lines, scuffs, chipped edges, a drain, under ordinary overhead strip lighting. The white laboratory void and the white cyclorama are the defaults this brief FAILS on, and so are the blue-lit sci-fi factory, the dark showroom with a single spotlight, and the neon-rimmed cyberpunk corridor. This should look like somewhere work actually happens. LIGHT: plain, overhead, unflattering, the light of a real building. If you want beauty, get it from the motion, not from the lighting rig. DISTANCE: change it inside the film. Start or end at a scale the eye cannot normally reach. FLOOR (hard, this is the material standard): a real text-to-video or image-to-video model, Seedance, Veo, Kling or Runway class. Frame interpolation of stills FAILS. A slideshow FAILS. A still with a Ken Burns push FAILS. Any on-screen type belongs to the shot, sharing its light and its grain, never bolted on after. Stock and archive footage FAILS. Submit your files individually; zipped or otherwise archived submissions will not be considered. If your model does not output 21:9 natively, deliver 16:9 and say so in the submission text; do not letterbox a 16:9 frame and call it 21:9. JUDGED: does it stop you, then is it true, then motion craft, light and sound. A beautiful film of the wrong fact loses to a rough film of the right one. NOT THIS: no robot uprising, no red glowing eyes, no Terminator styling, no robot-and-child handshake, no factory-of-the-future montage, no dancing robots, no humanoid staring meaningfully out of a window. No human faces, and no expressive or animated face on a machine; a blank head is fine. No on-screen countdown, no user interface overlay, no fake heads-up display with reticles and percentages. TEXT BUDGET: at most one line of type in the whole film, and only if it belongs to the shot. No credits, no captions, no lower thirds. NOTHING ABOUT HOW THE FILM WAS MADE GOES IN THE FILM: no process notes, no model or tool credit, no house credit, no timing or deadline text, and do not render this brief's section labels or the NOT THIS list as type. Credit your model in the submission text instead. DELIVER: one finished film at 21:9, 4 to 15 seconds, with sound, plus concept-note.md and sources.md. The FILM itself must be self contained, with no URLs, QR codes or watermarks burned into it; put your source links in sources.md. Name your model in the submission text.