# Housle

- **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:** [Clement Agbemafle](https://cerebralvalley.ai/u/Clwm)
- **GitHub:** https://github.com/Clem-Qhophi/BuildBrief
- **Demo video:** https://youtu.be/pd5TBJc6_qY
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4324

BuildBrief - Know what your AI built. From intent to proof, and from proof to understanding.

The problem: AI coding agents write code faster than any human can review it. You ask for X, the agent ships something adjacent, and nothing records what you actually approved, what the code proves, or what is still open. A finished-looking screen does not prove the rules behind it were implemented. Teams today either blindly trust the agent or re-review everything by hand. Existing reviewers scan code in isolation - none keep the persistent, evidence-backed link between human intent, implementation, correction, and team understanding.

What BuildBrief does: it is the consent-based, versioned system of record for software built with AI. You capture your exact request in plain language. BuildBrief extracts draft requirements, you clarify and explicitly approve them by name - drafts can never verify. It then checks that approved contract against revision-pinned code evidence and returns grounded verdicts: supported | partial | contradicted | not_established | not_applicable. Every verdict cites file:line at a pinned commit with snippet-hash check and names its evidence grade, so source (E1) is never confused with test (E2) or production proof. Code changes outside your contract are flagged as possibly unrequested with risk plus your disposition (explained/linked/approved/rejected) - a question, never an accusation, never auto-reverted. You get a bounded correction pack with scope, do-not-touch list, and verify-after steps. Only a new commit plus fresh recheck can close a finding; old verdicts go stale, history stays immutable in an append-only hash-chained ledger.

Why it is useful: a founder starts from an idea with no repo, approves a brief, and downloads the agent handoff. A developer sees the referral case fail honestly - 30-day expiry unproven plus a silent auth touch flagged critical - then watches the same acceptance command fail before the patch and pass after, flipping expiry to supported on fresh evidence only. A lead opens Understanding Studio for grounded explanations, role tours, and locate quizzes graded mechanically, while open answers wait for a human - comprehension is measured from real answers, never inferred. One engine everywhere: CLI, 15-tool MCP server delegating to tested code, Muse skill and hook handlers, and a local Next.js workbench on loopback :3103 reading the same ledger with no separate backend to drift.

## More from Meta Global AI Developer Hackathon

- [Keep the Light On](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4321)
- [Tulay](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4322)
- [Debitos](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4323)
- [Interaction Twin](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4325)
- [vivedex](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4326)
- [Oja Ledger](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4327)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4324. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt