This story has a hero, and it is not us.
In March, Glacis Technologies published OVERT — an open, royalty-free standard for proof at the AI runtime boundary. It says what a runtime has to be able to prove, not just document: that a check ran, what it decided, which rules applied, and that a stranger can verify it without trusting anyone. Joe Braidwood and the Glacis team wrote it, published it under a patent covenant anyone can build on, and then did something rarer: when we shipped Beacon without saying where the ideas came from, Joe wrote to us plainly, and we fixed it in the open. Beacon implements OVERT. The Foundation does not issue standards; we implement this one, teach it, and carry it to the people the market will not reach — the nonprofit, the county, the clinic, the classroom, the student.
The Fall 2026 Release is what carrying it looks like.
One practice, four levels, one hour each
Until now, the Foundation’s open-source projects were three downloads and a decision about which to install first. Now they are one practice with four levels, and each level is an hour or less:
- 100 · Begin. Nothing installed. Read the case of the airline chatbot whose promise bound the airline. Write the rule that would have caught it — in your own words, or as three lines of code; both work. Get a real, signed receipt from Beacon in your browser. Read it in a lens that is not your own. Bring it Thursday. Forty-five minutes.
- 200 · Retrofit. You have an agent, a script, a bot. Gate its three consequential actions. Put the rule in CI so that weakening it refuses the merge — the Knight Capital gate. Sixty minutes.
- 300 · Hold. Your agent touches people, money or rights. Name the human who releases a hold and how long it lives before it decays to no. Do the arithmetic that would have caught an insurer denying claims at 1.2 seconds each. State your coverage number, or state that you cannot count it yet. Sixty minutes today, a week of real traffic when the operate gate lands.
- 400 · Prove it to a stranger. Verify evidence you did not make, on a machine with no network. Write the gate that would have stopped a benefits scheme from issuing a single notice on an unvalidated formula. Host a Thursday.
Every level runs the same four gates — before an action, before a rule ships, while it runs, and after — and every level has a plain-words path and a code path to the same receipt. You do not need to code. You do not need to have read a regulation. You need an hour and a case.
The cases are real
The scenarios come from the Library’s 100 verified AI-harm cases — court rulings, regulator orders, inquiries — each carrying the gate that would have caught it. We chose the hardest ones on purpose: Air Canada, the dealership bot that agreed to sell a car for a dollar, Knight Capital, the Arup deepfake transfer, Cigna’s batch denials, Robodebt, the Dutch benefits scandal, the Post Office Horizon prosecutions. A practitioner who has only gated the easy cases is not ready.
The marks are the credential
When you finish a level and show the artifact to a person on Thursday, a host marks it, and the mark is a receipt on a public chain — verifiable offline with the same tool an auditor uses. Free to attempt, never sold. The levels map rung for rung to OVERT’s four assurance levels, and they describe your practice, never a system’s assurance; a system’s assurance is its own evidence, assessed the way the standard says.
Where it lives
Everything is on GitHub, which is the only source of truth: the one-page manifesto, the worksheets, the checklists, the release notes with the correct version of every engine and an honest table of what works today and what the Fall release lands. github.com/aigovops-foundation/practice — start with the manifesto, then open the worksheet for your level.
This is a release candidate. The founders decide the last open questions on 16 September, and the train leaves when one stranger finishes the level-100 worksheet in under an hour without us. If that stranger is you, we would like to know.
Ship safe AI — never unsafe AI. Get to yes. Stay at yes. Recover to yes. Prove it in ten minutes. Then bring it.