# Physics Proof

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Andrey Yunoshev](https://cerebralvalley.ai/u/yunoshev)
- **GitHub:** https://github.com/yunoshev/physics-proof
- **Demo video:** https://www.youtube.com/watch?v=xGzqiPTDuhI
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2814

Physics Proof answers one question about a video: is what it shows physically realistic, in the world the scene shows? Muse Spark describes the clip, then a second Spark call ("Wonder") asks what is worth proving and plans one to three checks: a free flight against the world's gravity, a size or speed against a sourced human range, a step rate against a natural clock. A Spark agent then measures the plan: SAM 3.1 segments the moving body and objects of known size, sourced web search gives their sizes, and the agent writes its own measurement code. A strict validator recomputes every number, checks that each quote came from the search and that every measured line lies on its object. Only then is a verdict drawn: realistic, not realistic, or undetermined with the reason. The output is an interactive player: the footage with masks, scale, trajectory fit and verdict drawn over it, the calculation beside it.

Demo: Apollo 16, the salute hop (NASA, public domain). Two sourced references (the backpack, 0.914 m, and the suited astronaut, 1.70 m) agree to 1.2 %. Both hops give g = 1.65 ± 0.10 m/s², 0.2σ from the Moon's 1.62. One run: 97 agent turns, $0.35. Our own test fake, the same clip sped up ×2.46, shows g = 9.58 ± 0.65, close to Earth's: the skeptic's trap. The frames show the Moon, so the verdict is "sped up ×2.43; at real speed it fits the Moon". Repeated twice on the final code: all four verdicts right.

Built during the hackathon: the only human-written part is 15 plain-English briefs (the owner's requests, turned into briefs by Muse Wonder). All code, 1,000+ tests and the player were generated by Muse Spark through Muse Code, one part at a time, with the full suite checking each generation. When a run went wrong, we changed the brief, not the code. A full rebuild costs about $0.33; one clip about $0.36 (Spark, SAM 3.1, search; all through the Meta Model API).

Honest limits: this is a working concept, not yet a tool for any clip. On a handheld street-jump clip, four runs have not reached a correct verdict yet: the agent took a hand on the car from small frames without checking its own masks, and the jumper's tucked legs need a wider error budget. Both are now general rules in the briefs; the regenerated code was running on that clip at the deadline. Many runs end "undetermined" because the validator refuses an unconfirmed number instead of giving a wrong answer.

## More from Meta Global AI Developer Hackathon

- [VRON](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2811)
- [Agent Rizz](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2812)
- [NeuralSyncLabs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2813)
- [Commonplace](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2815)
- [QTP](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2816)
- [Footfall](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2817)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2814. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt