# Djinn Council

- **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:** [Trent Taylor](https://cerebralvalley.ai/u/ChefTrent)
- **GitHub:** https://github.com/thecheftrent-coder/djinn-council
- **Demo video:** https://djinn-council.fly.dev/demo-video.mp4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/667

Djinn Council — a decision room for people who work alone.
Every AI is a genie: it grants your words exactly as said, not as meant.

Try it live: https://djinn-council.fly.dev/ (no sign-in, no keys needed). Watch the 3-minute demo: https://djinn-council.fly.dev/demo-video.mp4

Ask one AI a hard question and it hedges. It plays safe, lists pros and cons, and won't commit — because one model, asked once, has no reason to stick its neck out.

Djinn Council gives a solo decision-maker six ways to ask, on one ladder of effort:

- Rub the Lamp — one genie, one straight answer. About 15 seconds.
- The Stand — a straight question gets a straight answer, facts first. About 30 seconds.
- The Hearing — several models answer side by side; the differences are the information. About a minute.
- Cross-Examination — the models challenge each other's answers. About 1–2 minutes.
- The High Trial — the court researches the real world first, then decides. About 3–4 minutes.
- En Banc — the full court sits: an optional lawyer researches and briefs the case, counsel argues, the panel debates until it converges or hits the round cap, and the chairman rules. About 3–5 minutes.

Every verdict carries its confidence in plain sight — including when the honest answer is "low."

The test we ran on ourselves

We didn't take our own word for any of this. Before shipping, we ran the full court blind against one AI answering alone and against simple aggregation: 12 hard questions, protocol written in advance, scorer blind to which answer came from which system.

The council lost — 0 clear wins in 8 gate questions, and on one holdout question its synthesis failed outright and it never answered at all. We executed the fallback we had committed to in advance, then rebuilt the pipeline around the three failures the test exposed: synthesis reliability (a run must always land a verdict), decision quality (answers must commit), and wordiness. The demo video tells this story in the open, and docs/JUDGE_QA.md in the repo carries the full numbers.

Since the rebuild we have used it on real work, not demos. One straight question in The Stand set the strategy for an overnight Unity beta-bug hunt — new code, strange inputs, old projects opened in the new beta — and we followed it exactly. The next morning a full En Banc run, with counsel hired, gave our collaborator an audit-first repair plan for a 3D character clothing problem. Her verdict: it "actually helped a bunch."

Timeline, so nothing is vague: we found out about this hackathon and entered on Oct 18, 2026, and the entry was approved on Oct 20. The product idea was born that same week — a bridge between AIs in one room — so nothing about the concept predates this event.

Pre-existing (Oct 18–25, 2026): the concept and the intellectual framework — the architecture, the five-stage Forge design, mode playbooks, the blind-shootout protocol, and the kickoff runbook — plus a repo scaffold (first commit Oct 22) holding those design documents and a starter code skeleton: an orchestrator draft, a stub Meta adapter, and a frontend shell. It had never been run. There was no working software.

Built during the hackathon (Oct 26 – Nov 4, 2026): everything that works — the first verified end-to-end run (Oct 26) and the whole product as Djinn Council: the name and genie framing; live Meta, OpenAI, and OpenRouter adapters with Muse Spark in the chairman's seat; The Stand and The High Trial; spend controls, the failure-injection suite, and the shootout harness; demo mode with server-side keys; the council scene; the Day-3 blind shootout itself; and the post-verdict rebuild — Stage-0 counsel, the convergence debate, and the synthesis reliability floor.

The hardest part was the movie behind the button

The scene that plays while the council thinks took more than sixty generated takes across three days. AI video does not remember: the table changed shape between takes, judges picked up pens they were never given, finished symbols lit themselves again.

So we stopped trusting single takes — the same move the product makes. We built a director's harness: locked reference pictures re-fed into every take as memory, every take checked in motion (not in stills), each new piece chained from a checked frame of the one before. Only the clerk writes. A symbol lights when the next one starts, then dims. Nobody sells that discipline — we built it, for the film and for the answers alike.

Extra credit — how the council scene was filmed (38-second bonus, narrated): https://djinn-council.fly.dev/animation-bonus.mp4

## More from Meta Global AI Developer Hackathon

- [Temi AI by NuvoAI.](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/664)
- [FounderFlo Ops](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/665)
- [Phoenixon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/666)
- [Momar SOZO](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/668)
- [Cortex AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/669)
- [Ontology-Physiology Lab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/670)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/667. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt