# Xinogi

- **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:** [Ryotaro H](https://cerebralvalley.ai/u/xinogi)
- **GitHub:** https://github.com/xinogi/gravity-check
- **Demo video:** https://youtu.be/ceGhpBfIvh4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1457

AI video models forget physics. A football stops in mid-air and never lands, and a woman told to carry three oranges ends up holding four. People find these breaks after they post, because nobody can watch every frame of every take.

Gravity Check reads an AI video and the prompt that made it, then shows where the video breaks physics.

Muse Glimmer reads the numbers the prompt promises and names objects in each frame that nothing seems to hold up. SAM 3.1 cuts out those objects and the hands and arms. Plain geometry on the masks makes the call: is it touching a holder, resting on a held pile, moving between frames, or more than the prompt asked for?

The report gives the time, the object in red, the reason, and a prompt to paste back into the generator for the next take. It all runs on one PC with a 12 GB RTX 4070. The clip never leaves the machine.

It is a checker, not a guarantee, and I tested it blind, four times. I asked Muse for clips that break physics and never showed the checker those prompts. Claude, separate from the checker, labelled the frames and scored the results, and I have not confirmed every label by hand.

- First test, 15 clips: caught 7 of 10 breaks, with false alarms on 2 of 3 normal clips.
- Second test, 10 normal clips: false alarms on 4.
- Third test, 15 new clips, after adding Glimmer's second look: caught 5 of 5 breaks, with false alarms on 3 of 10 normal clips.
- Fourth test, 10 clips of stacks, on the current code: caught 3 of 3 breaks, with a false alarm on 1 of 7 normal clips (bananas hanging at a market stall).

Two of the third test's three false alarms lasted a single frame, so one-frame findings are now dropped. On those same clips that leaves 1 of 10, but that number is not a blind result. The fourth test asked one question: does a pile, or something carried by a body, look like it is floating? Balls stacked on a palm stayed clean because support passes along the pile, and books on a head stayed clean after Glimmer's second look. The geometry itself only knows hands and arms, so a table, a lap or a head rests on Glimmer's judgement. The breaks in the last two tests were asked for on purpose.

## More from Meta Global AI Developer Hackathon

- [GridFork](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1454)
- [Semantic Release Auditor](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1455)
- [Sachitha Madhuranga](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1456)
- [Ajay Upadhyay](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1458)
- [HocBong Hunter AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1459)
- [Jinboost](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1460)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1457. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt