# CivicScope

- **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:** [Luke Warm](https://cerebralvalley.ai/u/Lucasio)
- **GitHub:** https://github.com/Treystu/musehackathon
- **Demo video:** https://youtu.be/Rh1L3NkuonA
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4797

CivicScope makes US housing and neighborhood data explorable by anyone.
Every figure is sourced, and every figure carries its margin of error.
It's live at https://civicscope-muse-portal.pages.dev — one surface,
zero setup. A shared judging key is built into the deployment, so Ask and
Explain work on first load and judges never handle keys.
The problem: US Census and survey data (ACS / TIGERweb, CDC PLACES, NCES
EDGE) can answer the housing questions people actually ask — where is rent
lower, which places meet my criteria — but only through specialist tools.
General AI answers just invent the numbers.
CivicScope's deterministic data layer covers 33,791 ZIP areas. The
cheapest median rent in the dataset is Corbin, KY 40771 at $99/mo; the
dataset median rent is $1,286/mo.
For this hackathon we built a Muse-powered layer on top of that data:
natural-language Ask, validated Explain, a briefing pack, and
agent-assisted key onboarding. It runs under one strict rule — the model
proposes intent and prose, deterministic code owns every number, and a
validator gates anything the model produces before it can render.
Ask turns a plain question ("Compare rent in 78701 and 60601", "Hawaii ZIP
codes with rent under $2,000") into a structured, editable query that runs
against the real data. Explain reads the retrieved figures back in plain
language — and any sentence containing a number that was not retrieved is
withheld in code before it can display.
Safety is mechanical, not prompted: allowlisted metrics and filters,
validated output, and an aggregate-only usage meter that counts and
reports only. The meter is accounting only — it never blocks, gates,
delays, or fails a user request.
The layer is evaluated on a versioned 91-case golden corpus with a 7-case
safety subset. Measured live on the re-frozen build (`94f796e`) on
2026-11-04 (direct surface), it scored exactIntent 96.7% (88/91) and
safety 100% (7/7) — both bars met.
What predates the hackathon is the deterministic data layer and its
metric registry, retained unchanged in substance; the Muse layer on top
of it was built during the contest.

## More from Meta Global AI Developer Hackathon

- [Venox](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4794)
- [LabSpotter](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4795)
- [Magic](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4796)
- [Jovinho](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4798)
- [Data Alchemist](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4799)
- [Atul Khatri](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4800)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4797. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt