Vacant

The problem

Plenty of people can vibe code. The most common failure of AI agents isn't failing to write code — it's saying "done" when it isn't. You either babysit it or trust it. And "I tested it" at handover leaves no trace of what was actually delivered.

What we measured

What I built

On a team of six, I owned dev, experiments, and theory: I led Vacant's design and implementation, planned and ran the pre-registered experiments, and built the exhibition's digital-twin system and the website.

Acceptance gate
A broker wrapped around the AI agent. "Done" doesn't count until your acceptance check passes; failures are recorded as rejected with a non-zero exit code.
Signed receipt chain
Who did what, and what was checked — each cell locks the previous one. Change one character and the chain breaks right there. Verifiable offline, recomputable with zero API calls.
The vacant menu
Keep the AI you already use. Type vacant and pick an agent on your machine — Claude Code, Codex, OpenCode, pi.
Pre-registered experiments
Criteria written before the run. 32 lab notes and a 58-claim evidence ledger, with full numbers and denominators.
Exhibition digital twins
The clay residents are digital twins that appear in the physical exhibition.
This website
A mechanism simulation that hashes receipts live in the browser, so non-programmers can see what was verified.

What it won't do

• ✕Certify an AI as reliable — a grade belongs to one run, not to a model.
• ✕Make the AI smarter — 30 held back, 0 rescued by retrying.
• ✕Grade work that can't be written as a test — no runnable acceptance, no referee.

What I learned

An overturned claim matters as much as an upheld one. 14 of our 58 claims were overturned by our own experiments, and they all stay on the ledger. Retrying rescued 0 of 30 — agents don't get smarter by trying again. What you can do is make "it didn't pass" visible, recorded, and accountable.