# Stardive

- **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:** [Jiaye Tey](https://cerebralvalley.ai/u/icedamericao)
- **GitHub:** https://github.com/JyTey2004/muse-audit
- **Demo video:** https://www.stardive.xyz/hackathon/
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4406

As agents become more autonomous, we need better tools for assurance of agentic work. When an agent says a task is done, users need a way to inspect what was checked, against which requirements, and whether that evidence still matches the delivered result. Stardive is a step toward that assurance layer.

Your AI says it’s done. Stardive makes it prove it. Stardive connects a saved request, the version of a file that was checked, and the file delivered to the user through an inspectable completion receipt. Checks keep PASS, FAIL and UNKNOWN separate rather than treating an AI’s completion claim as proof.

The demo page presents a recorded consumer Muse workflow creating a welcome-note text file, checking required text and a 20-word limit, and delivering a content-bound receipt. Its published task record reports two PASS checks, a nine-word, 48-byte result and SELECTED_CHECKS_PASS. This is a narrow deterministic text-and-word-count demonstration, not general semantic verification or proof that every requirement was captured. The 43-second video uses actual native screenshots with waiting condensed; it is not a continuous screen recording.

The public muse-audit repository documents the earlier offline MCP prototype: four read-only audit/repair-planning tools, a shared verification kernel, synthetic XLSX budget and DOCX study-summary examples, and HTML/JSON/Markdown reports. It checks explicit contracts, protects declared commitments, detects stale evidence and applies reviewed repairs only through a separate CLI into a new copy. The public repository and newer recorded consumer workflow are distinct evidence sources. The published demo record does not export model identities or a complete hidden action trace; no specific Meta model or separate-model semantic judge is claimed.

## More from Meta Global AI Developer Hackathon

- [The Magic Table](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4403)
- [Zenith](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4404)
- [Innovian](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4405)
- [Ishan's Team](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4407)
- [David Castro - Agent Guard](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4408)
- [dark%hunter](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4409)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4406. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt