Aleph Hackathon

The problem Most teams submit a project nobody reviewed. The obvious fix is to paste it into ChatGPT — but you don't, because the idea isn't public yet. So you find out what's broken from a judge, after the deadline.

What JUD does Pick the hackathon, the track, and what you're pasting. Drop in your pitch, your README, or your code. A model running on your own laptop returns a report:

• Aura, Laura, or No judgment
• A score, and the rubric broken down dimension by dimension
• The exact hard fail you hit, in plain language
• Quotes pulled from your own text as evidence
• A one-line roast, a one-line fix, and the gaps it could not verify

JUD is the product. Aura or Laura is the verdict.

Why it has to be local

No API keys. No accounts. No cloud. The inference server binds to 127.0.0.1 and the browser never talks to it directly — Next proxies /api/* to loopback. Nothing you paste leaves the machine.

That isn't a privacy checkbox. It's the only reason you'd paste an unpublished project into it at all.

The rubric is a file, not a prompt

The criteria live in packs/aleph-2026.json: tracks, hard fails, and scoring dimensions.

• Hard fails are regex, evaluated in code before and after the model runs. Cloud inference on the QVAC track, a bolted-on wallet on WDK, a missing pear:// on Pears, or no demo at all — those force Laura regardless of what the model thinks.
• Hard passes work the same way in reverse.
• The rubric breakdown is computed in code. The 1B model extracts and explains; it does not invent the fail.
• Below the pack's minimum length, JUD abstains and never calls the model. It won't score vibe.

Same pitch, same verdict, every time.

Clone it for another event

Copy packs/_template.json, fill in tracks, hardFails, and dimensions, and JUD is judging a different hackathon. No code changes. Aleph 2026 is just the first pack.

Stack

• Next.js 16 + React 19 + Tailwind, shadcn/ui
• @qvac/sdk running Llama 3.2 1B Instruct Q4_0, JSON-schema constrained output
• Node 22.14+, ~773 MB model cache, no GPU required
• UI and generated report both available in English and Spanish

Run it

npm install
npm run dev

Open http://127.0.0.1:43123. First launch downloads the model once; after that it works offline.

Track

Aleph Hackathon 2026 — QVAC Track 1: local agents for operations work. Messy text judged against a spec, with visible reasoning and flagged uncertainty. Here the document is a pitch and the spec is the sponsor's rubric. The same architecture handles any document-versus-rules review where the document can't be uploaded anywhere.