# RECALLZERO

- **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:** [Guelmbaye Firmin Djikoloum](https://cerebralvalley.ai/u/guelmbaye)
- **GitHub:** https://github.com/guelmbaye/recallzero
- **Demo video:** https://youtu.be/77MjHg4UEaM
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1787

RECALLZERO — One More Photo. A Different Decision.

What happens when an AI must decide not just what it sees, but whether it has enough evidence to act?

RECALLZERO explores a new capability for multimodal AI: turning incomplete, conflicting, and potentially misleading observations of physical objects into explainable, evidence-aware decisions.

Powered by Meta Muse Spark 1.3, RECALLZERO examines product photographs, identifies missing information, compares markings across multiple views, and recommends whether to request additional evidence, allow a narrowly defined visual-consistency check to pass, or hold a case for human review.

In our working demonstration, one photograph shows a RYOBI battery marked 6Ah. Muse requests another view because the model code is missing. When a second photograph reveals RB18L40, Muse identifies a suspected discrepancy and recommends holding the case for review. An unverified claim that the item is genuine does not reverse that decision.

We built six pre-sealed evaluation conditions using publicly sourced photographs. Muse matched all six predefined decisions, compared with five of six for a conservative rules baseline using human-transcribed markings.

The system includes SHA-256 evidence sealing, blind model inputs, archived reasoning outputs, a verifiable replay interface, and a simulated policy controller.

This is an exploratory physical-world reasoning prototype—not a counterfeit certification system or a live checkout integration.

Our frontier: machines that recognize when evidence is insufficient, revise decisions when new observations arrive, and preserve an auditable explanation of why.

## More from Meta Global AI Developer Hackathon

- [DecideOnce](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1784)
- [SuperHornet](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1785)
- [Endovascular ToolKit](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1786)
- [WellArtDev](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1788)
- [Ilma](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1789)
- [OfferTest AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1790)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1787. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt