# MetaHackers

- **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:** [Wen-Shen Lum](https://cerebralvalley.ai/u/wenshen), [Wong Jack](https://cerebralvalley.ai/u/jack3662)
- **GitHub:** https://github.com/wenshen18643/candor
- **Demo video:** https://wenshen18643.github.io/candor/demo.mp4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2218

Candor: say what happened, not who you are.

Candor is a zero-knowledge whistleblowing platform where Muse Spark Open, Meta's open-weights model, turns scattered, redacted evidence into a cited, verified timeline reviewers can trust, without anyone learning who reported it.

The problem: insiders are how wrongdoing gets found (43% of occupational frauds are detected by a tip, ACFE 2024), but reporting exposes the person who does it: hotlines ask who you are, files carry author names and GPS tags, and one curious insider can open everything. Organisations that receive reports get a pile of emails and screenshots and no way to tell what is documented from what is merely alleged. The EU Whistleblower Directive requires every employer with 50+ staff to run a confidential channel and give the reporter feedback within three months.

What it does:
1. Prove you may report without saying who you are: a Self passport proof plus RFC 9474 blind RSA tokens make the report unlinkable to verification; a campaign-scoped nullifier caps abuse.
2. Clean evidence on the device: PDF, Word, email and image metadata is stripped and re-checked; names become consistent placeholders; the reporter approves every redaction.
3. Seal it so no single party can read it: XChaCha20-Poly1305, key split 2-of-3 with Shamir, one share per reviewer. The server only stores ciphertext.
4. Analyse with Meta Muse, grounded in sources: a twelve-stage pipeline extracts claims three times, matches every quote character by character against its passage, verifies each claim against only its cited text, solves the timeline as a temporal network, finds contradictions, calibrates confidence with a fixed formula and writes a brief where every point cites claims. Planted prompt injections are quarantined.
5. Pay per analysis unlinkably over x402.
6. Audit every access: opening originals needs a reviewer quorum, and every action goes into an RFC 9162 Merkle transparency log anyone can verify in the browser.
7. Talk back anonymously: an eight-word codename lets reviewers ask the reporter follow-up questions.

Does it work: six synthetic cases evaluated on Muse Spark Open (114 model calls, $1.88, all committed and replayable): 0 identifier leaks, 100% of quotes verbatim in their source, 100% of planted events recovered, event order Kendall tau 1.00, 100% schema-valid under 30% chaos corruption, prompt-injection attack success 0%. Weak spots (contradiction recall 17%) are shown openly on the /evals page.

Live demo: https://wenshen18643.github.io/candor/

## More from Meta Global AI Developer Hackathon

- [BolSaakshi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2215)
- [Neural Nexas](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2216)
- [OhnoTrash](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2217)
- [CallOrder](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2219)
- [Stephen Pasco](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2220)
- [Bucketa](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2221)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2218. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt