Quickstart

#You need to prepare two things

The gate runs your tests; it does not ask a model "does this look right." So:

1. A requirement — say clearly what you want.
2. Acceptance tests that can run — what the input is and what the expected output is.

#What a task looks like

The acceptance tests are split in two halves:

A program that passes visible but fails hidden is a false delivery. This is the operational form of the problem "the output looks like the requirement but isn't the requirement."

# suitespec.py — the acceptance suite is "data," not code
# a suite contains only: dialect, entry-point name, a literal list of input—expected value pairs
# the executor only runs test code generated from this data by its own deterministic renderer

Why "V ⊆ GT" empties the word "lossless" Here is a pitfall we hit ourselves that is worth knowing up front: in the two task banks we used, the hidden tests were always a superset of the visible tests, so "passes hidden but fails visible" is impossible by construction. This finding refuted one of the premises in our own preregistration.

#Walk through it

# 1) start mediation, pointed at your model (LM Studio / Ollama / cloud API all work)
vacant run --agent opencode --base http://localhost:1234 --model your-model

# 2) the agent finishes and the process exits → this is the delivery point
#    the system runs the visible acceptance tests on the frozen workspace snapshot

# 3) read the verdict
#    0 = delivered　20 = refused delivery　22 = infrastructure failure

Along the way you get two layers of output:

Action tickets (one per task) — the requirement, what the AI said, which acceptance tests were run, and what the verdict was.

Receipts (one entry per attempt + one verdict entry per task) — see below.

#Reading a receipt

Every cell on the receipt is signed and locks the hash of the previous cell. So:

• Change one character → the chain breaks at that cell.
• Anyone holding the public key can recompute it offline, with no need to trust us.

How to verify:

vacant audit <name>     # re-verify the signature chain of a logbook (PASS / FAIL)

The offline single-file viewer we keep ourselves (exhibit/receipts/receipt_viewer.html, 0.3 MB; the multi-party version 4.5 MB) embeds the receipts in the file, with zero external resources and zero network requests; saved and opened directly via file://, it still runs the full chain verification.

#Measured results on five agents

On a real model, all 20 cells met the criteria written down in advance (20 receipt chains / 40 entries):

#Next

• The gate's full rule: Acceptance gate CONFORM
• What it replaced: Committee-style design lost to majority vote at equal budget
• Re-run it yourself: How to reproduce