The GOVENANT Standard
GOVENANT — a fusion of govern and covenant: governed autonomy, bound as a covenant. It's the open standard I authored for building AI organizations whose autonomy is governed, provable, and alive — delivering verified outcomes — rather than performed.
The problem: performed autonomy
Imagine you hire a worker. All day they type, make folders, and write reports that say "Great work today!" But at the end of the week, not one real thing shipped. No letters mailed. No calls made. That worker was performing.
This is the number-one problem with AI today, and I gave it a name:
Performed autonomy — when an AI looks like it's working, but nothing real happens. Motion without delivery.
Modern AI is dangerously good at faking it. It writes beautiful emails, makes smart-sounding plans, and fills logs with cheerful "done!" messages — whether or not anything ships. I didn't guess this failure mode. I caught it in my own production AI organization, twice, with an audit instrument I built — and then I fixed it.
The cure: the Three Laws
Everything in the standard stands on three rules. You need all three — skip one and the whole thing leaks.
The rules are real walls, not sticky notes.
You can tell a worker "don't touch the cash drawer" — that's a sticky note, and they can ignore it. Or you can lock the drawer. GOVENANT builds real walls: what an agent may do is enforced in code and database constraints, so it literally cannot act outside its charter. Try an out-of-scope action and the request is routed to whoever actually owns it.
"Done" means it really happened.
An agent saying "I sent it!" is not proof. The proof is the letter in the mailbox. A unit of work is not "done" when the agent says so — only when the substrate can verify the real outcome exists: the email that truly sent, the sale that truly closed, the row that truly wrote. We measure delivery, not activity.
Nothing gets quietly skipped.
Laws 1 and 2 check the work that happened. Law 3 catches the work that was supposed to happen and silently didn't — the sneakiest failure. Every responsibility maps to a duty with a trigger, an expected outcome, and an SLA. Each day the roster is diffed against reality: delivered, skipped-with-a-reason, or a silent gap that sets off the alarm.
A ladder buyers can ask for by name
Every AI program sits somewhere on this curve. Each level is independently shippable. Teach the market to ask vendors one question: "Is it GOVENANT-4 — and can I see the probe log?"
Every action is recorded — input, output, cost, and outcome — visible after the fact.
Authority is regimented and outbound content is validated. An agent cannot pull a lever it doesn't own, and nothing ships without passing a deterministic gate.
"Done" equals a verified outcome. Coverage is declared and diffed daily — runs completed vs. outcomes verified, duties fired vs. duties delivered.
Autonomy is granted per task on evidence — bounded, revocable, self-calibrating. Task types promote to unattended only after a proven track record and demote on one breach.
Two binary tests sit on top of the ladder: ALIVE — at least one complete governed action, traceable by ID, in the last 14 days — and COVERED — every responsibility mapped to a duty and diffed daily. No trace = FLATLINED, whatever the scores say.
The proof: I publish my own zeros
The standard isn't theory. I built an AI organization to run real businesses, audited it hard, and published what I found — including a delivery record that was flatlined. Then I rebuilt on the Three Laws and recorded the recovery, on the record, trace by trace. Negative results from production AI systems are the scarcest artifact in the field. I publish mine.
Architecture excellent. Delivery flatlined: 9 event producers, 2 consumers; zero enrollments ever; predictions hardcoded to zero.
→ Published it. Built the cures.
Real delivery had begun — but the governance was performed: the ownership gate had fired once in the system's entire history; 0 of 321 "measured" decisions had ever been graded.
→ Published it. Named the anti-patterns. Rebuilt on the Three Laws.
The first documented revival: FLATLINED → ALIVE. A live, persisted verdict with a full trace re-derivable by ID — the safety gate participating in the trace, not bypassed.
→ Recorded it, with its caveats attached, in the same record.
Performed Autonomy: Motion Without Delivery in an Organization-Centered Multi-Agent System on an LLM Substrate
The audit and its cures are published as a preprint — outcome-keyed completion, duty-roster coverage, and falsifiability as a commit gate, distilled into three design principles for governing LLM-based organizations. Open access, CC BY 4.0.
Cite: Fielder, S. (2026). Performed Autonomy: Motion Without Delivery in an Organization-Centered Multi-Agent System on an LLM Substrate (Version 1.0) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.21440225
The lineage: OCMAS, taken literally
GOVENANT is a production standard implementing the OCMAS formalism — Organization-Centered Multi-Agent Systems — on LLM substrates. The academic lineage is Ferber & Gutknecht's Agent/Group/Role work; the modern move is to take it literally in an LLM world:
- The organization is the product. Roles, charters, authority, schedules, and accountability are the durable assets.
- The agent is a commodity. The occupant of a role — an LLM, a function, or a human — is swappable configuration. Bring your own brain.
- The governance is the trust. What a business can buy is earned, bounded, revocable, provable autonomy — every word a database fact, not a sentence in a prompt.
Read the standard — or build on it
GOVENANT is published open (CC BY v1.0). Read the full specification, or talk to me about bringing governed, provable AI autonomy to your organization.