Vacant
An accountability layer for AI
GitHub
The moment the AI says it's finished, Vacant checks it once against your standard and leaves a receipt you can recompute yourself.
9/11 Three AI assistants did 10 tasks each. Of the 11 that were stopped at the gate, 9 were ones the assistant itself had declared "finished."
Every clay resident is a digital twin that appears at the exhibition.
Mechanism simulation: the visuals are illustrative, and the AI's replies are three prerecorded scripts. The receipt hashes are computed live by your browser.
Each point comes with the number we actually measured.
01
When the AI says it's done, Vacant first runs your acceptance check. When you come back, you only need to look: it passed, or it was stopped at the gate.
9/11 Of the 11 that were stopped at the gate, 9 were ones the AI itself had declared finished.pi, Claude Code, and OpenCode, 10 tasks each. Raw data
02
The receipt records it as a refused delivery, and the exit code is not 0, so your pipeline sees that and won't accept it.
30 stopped Across the 54 cells of the exhibit, the 30 failing pieces that would have shipped without this layer were all stopped at the gate, and each left a receipt.But retrying rescued 0 of them (in 27 of them, the task deliberately hid the function name). Under the default settings, refusal is a marker, not a wall; to actually stop the files, you have to run in a sandbox. Raw data
03
Who did it and what was checked: each cell locks the one before it. Change one character, and the chain breaks at that cell.
9,841 entries 194 real receipt chains, all of which passed independent signature verification. Verify it yourself
This is the receipt for the ticket you just saw (mechanism simulation). Click a cell and change one character.
04
Vacant wraps around it; it is not another AI. Type vacant and pick an assistant you already have on your computer.
5 assistants Hooked up to the gate on the same machine, 20/20 cells met the criterion written in advance.But they judge differently: three of them agreed on only 4/10 tasks.
05
The same set of 120 tasks, repeated five times. "Revise from the failure message" came out slightly above "rewrite" all five times, but none of them survived statistical correction — not measured, which is not the same as no difference.
"The whole thing rests on 'requirements can be compiled into executable acceptance tests.' … Where a requirement can't be run, this mechanism has no free referee. It degrades into 'ask a model' — and that is exactly the thing that measured badly."整件事建立在「需求可以被編譯成可執行的驗收測資」。……需求跑不起來的場合，這個機制沒有免費的裁判，會退化成「問一個模型」，而那正是量出來很差的東西。The precondition of every effectiveness number (R440P §五-1; translation, original Chinese above)
This section is a hope, not a result
Only in a run that is locked inside the wall does every call go through the gate.
A level belongs to one run, not to an AI.
30 stopped at the gate, 0 rescued by retrying.
Without an acceptance check that runs automatically, there is no referee.
Go deeper
How to install it, and which assistants it can hook up to, is all on GitHub.
Every name is tied to a piece of work you can ask questions about. Click a name to see what that person is responsible for.
陳建安PM · Progress and quality controlOversees and manages every member's progress and controls the quality of each member's work; relays the curation team's announcements so the group receives exhibition news. Reviews the developer logs and meeting minutes, so that mistakes don't leave us unable to tell what happened at the time.
Department of Information and Communications, Shih Hsin University 2026 capstone project · On-site exhibition
