MAiFactory governed by the MAI-Meta Foundation
A factory floor for multi-AI work.
MAiFactory assembles specialized AI roles to design, build, validate and deploy work under the MAI-Meta Foundation — a constitution the agents have no standing to amend.
Most AI systems optimize for speed. The harder question is whether, six months later, the work remains attributable, governed and reconstructable.
Not only the outcome — the path to it. Roles, evidence, disagreements, approvals, all still readable after the session ends.
Governance decisions require different provider families. No provider family validates its own work.
Unreliable work is marked inadmissible, its trace preserved, and rebuilt from clean authority.
- Teams and specialist roles
- Projects and work environments
- Validation and deployment workflows
- Evidence and recovery records
01 — The Factory in operation
Four roles, and the things each one may not do.
Select a role to see its prohibition
The prohibitions are the point — they make this a chain rather than four assistants working on the same folder. The Foundation defines the rules, MAiFactory executes them, and the proof cases show the resulting work.
The standing roles.
The chain runs once per package · these never stop
The four workflow roles complete and hand on. These do not complete. They run for as long as the Factory is running, and they share a single prohibition: they detect, and they do not repair.
Correct the entry. The correction is directed to the role that recorded it, and revalidated before it stands.
Resolve what it found, or clear a condition it raised itself.
Add substance to what it routes, or gain interpretive authority by sitting between two roles.
Seat two directors from a single provider family.
Direct the work, or clear it.
The correction route — fixed, not case by case
DETECT RAISE DECIDE REPAIR RE-CLEAR ────── ───── ────── ────── ──────── Recorder ─┐ ├──▶ Orchestration ──▶ Operator ──▶ originating ──▶ Validator ──▶ commit Sentinel ─┘ Assistant (Principal) role / push
The role that finds a fault never fixes it, and the role that fixes it never clears it. Nothing reaches the repository without a fresh validation pass.
To date the Recorder has never found an incorrect entry — the Validator catches them first. A control that has not yet fired because an earlier control caught everything is defence in depth, not an idle role. Governance configuration is maintained by a Custodian, which operates on the Factory itself rather than on the work it produces, and so does not appear on a project team.
02 — How a Factory team works
A team, a work surface, and a record of everything on it.
- Governance and control structure
- Team configuration
- Records and evidence
- Manufacturing → project repository clone (the production surface)
An operator launches a Factory with a named team. Its work surface carries a manufacturing folder holding a git-protected repository — where the team specs, designs, validates, builds and tests.
Four workflow roles form the chain; five standing roles run continuously alongside it. Both are set out in section 01. A team is that roster instantiated against one project — the roles are constant, the work surface is not.
Nothing moves on assertion. The next role reviews against the constitution and the evidence cited, and a rejection returns with a defect classification.
Every approval binds to a versioned artifact and its content hash. Nothing is edited in place, and decisions are signed into a chain-verified trust ledger.
Every role reads the repository. One role writes to it. Nothing is committed until the Validator has cleared it — the Builder holds commit authority but not the decision to exercise it. Commits land on package completion, so commit granularity matches package granularity. Publication is a further authority the Principal exercises on his own timing: immediately after a patch, or once several packages have accumulated. Three roles, three moments — and the role that writes to the record is never the role that judged it fit to be written.
Ambiguity that matters stops the work and surfaces to the Principal. Failing closed is the default, not the exception.
The Principal can halt, redirect or override at any point — carrying each handoff personally, confirming proposed actions, or setting the constraints an autonomous run executes inside.
The roster is expanding beyond engineering into UX design, marketing and PR and further advisory roles, each bound to the same constitution on the way in. Adding a role never touches the Foundation.
03 — What governs the Factory
The MAI-Meta Foundation.
A constitution, not a configuration file
The Foundation holds the law: what may be done, by whom, and how compliance is proven. MAiFactory holds the work: the team, the project, the repository. Agents reside in the Factory and cannot amend the law they operate under.
04 — Operating modes
The same law, with and without a human in the loop.
Human-orchestrated
The Principal carries every message and authorizes every transition. The reference implementation: if the law cannot be executed by hand, it cannot be automated honestly.
Software-orchestrated
A governance-aware orchestrator routes work through the same chain, enforces separation of duties, checks provider-family independence, and signs every decision into an HMAC-chained ledger.
- Observation Recommends only. Nothing executes.
- Supervised Proposed actions wait for confirmation.
- Autonomous Executes inside constitutional constraints, logged end to end.
One team takes a project end to end inside its Factory.
The automated development factory: orchestration, not the operator, carries the work between roles.
05 — Proof cases
Two problem classes, one constitution.
Operational diagnosis · governed data extraction and transformation
ADAM ML and the Engineering Tool
Analytical and runtime complexity carried through role-separated validation to a production-aligned package.
Read the case Case 02DPT-Workbench
Structure-preserving pseudonymization with downstream fidelity intact, and a governed recovery after evidence contamination.
Read the case Case 03DoorSet-CSV-Extract
Governed extraction that reconstructs topology at the source and fails closed on schema drift.
Read the caseMAiFactory is the operating system. The MAI-Meta Foundation is its constitutional authority.
Or write directly — dennis@maiframeworks.com