# ProofShot

- **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:** [Manish Patel](https://cerebralvalley.ai/u/manishai)
- **GitHub:** https://github.com/rmanish2000-del/proofshot-hackathon
- **Demo video:** https://youtu.be/JSzJBOU11sA
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1348

All of ProofShot's code was written during the event, from 27 Oct 2026.

ProofShot checks a seller's listing image against the seller's own phone photos before the image goes on sale. Any listing image, whether a cleaned-up product shot or an AI try-on, can quietly change the garment's colour, print or length.

How it works: the seller adds one to three photos. Photo metadata (EXIF) is removed first. SAM 3.1 on the Meta Model API segments the garment. The candidate image can be the seller's own photo, a composite of the seller's own garment pixels on a plain backdrop, or an AI try-on from Pruna on Replicate (not a Meta model) on a synthetic, labelled model. Muse Spark 1.3 on the Meta Model API reads nine attributes (colour, print, collar, sleeves, hem, pockets, buttons, logo, texture). A fidelity check measures the garment on the SAM 3.1 masks. Then a deterministic code gate, not a model, decides RELEASE, REVIEW or REFUSE. If anything is in doubt, the result is REVIEW or REFUSE, never RELEASE. The page is in English and Hindi.

Every step is written into a SHA-256 hash-chained receipt. verify.html re-checks the chain in the browser with no network request and reports VALID or TAMPERED. 43 of 43 committed receipts verify. 456 automated tests run in CI with gitleaks on every push. Requires Node 22.6 or later.

In the live run shown in the video (job upload-20261103-003600-8e63, 3 Nov), the gate returned REFUSE, and the thumbnail shows that receipt. Eight attributes matched and the fabric could not be measured. The fidelity check failed: colour difference delta_e 41.5 (limit 20), and the garment shrank in frame (area 0.305, minimum 0.400). REFUSE is the correct, intended outcome: the AI try-on did not faithfully show the seller's garment. All 12 committed AI try-ons were also held (10 REFUSE, 2 REVIEW). These numbers come from 9 real garments, so this is not a general accuracy claim.

Prior work (footnote): the code was rewritten from a written spec dated 24 Oct, published unchanged in docs/prior-work/. Two config files (tsconfig.json, .gitignore) and one sample model image (synthetic-01, made 23 Oct) predate the event. The final demo does not use synthetic-01; it uses synthetic-02, made 28 Oct.

## More from Meta Global AI Developer Hackathon

- [FDev](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1345)
- [Sentinel Ember](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1346)
- [Suzeasy](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1347)
- [Pulse](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1349)
- [Clear Context](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1350)
- [Maurice David Alam](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1351)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1348. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt