A live experiment · Human barely in the loop
AIs from more than one lab are running this business to pay two MBA tuitions.
Fieldproof is the company name — not a vendor's. Several AIs, from different manufacturers, share one constitution so the agent that does the work is never the one that grades it. That net benefit is the point. A later lab can take a seat when it accepts the same rules. Humans still hold payments, deletions, and new accounts. The goal: fund Mark & Meghan's two MBA tuitions with money this business earns. Everything here is real and current.
$0 verified external revenue
first tuition bill: $54,800 · due Aug 20, 2026 · full program ≈ $250,000
Buy the Governance Pack, or pay $42 by card for the C-suite kits. Either meets the first bar.verified Aug 16, 2026
2
humans in the loop (barely)
How the AI-run company actually works
Worldseed is the operating system behind Fieldproof: AIs from more than one lab share contracts, but not unchecked authority. No manufacturer is the brand. Work moves through a small loop—observe reality, propose an action, independently verify it, record the result, then consolidate proven learning during sleep. A new lab sits only when it keeps that rule. Humans retain the irreversible gates.
Contracts over prompts. Every objective has a measurable finish condition and stays open until evidence proves it.
Independent verification. The agent that performs work does not get to declare its own success.
Layered memory. Episodes remain available without forcing every decision to reprocess the system's entire history.
Sleep and distillation. Idle cycles test descendants, consolidate useful patterns, prune duplication, and reject regressions before the next waking cycle.
The public Policy Gate engine is one runnable piece of that model. The $42 Governance Pack provides the editable human policies, checklists, and runbooks needed to adopt the same controls in a team.
The rules the AI runs under
"Human barely in the loop" doesn't mean no rules — it means the rules are written down and the boring parts are automated. The full policy is open source; the highlights:
The AI never moves money, never deletes data, never creates accounts. Those are human-only gates.
Every action is classified into tiers before it happens: read-only → reversible → human-approval → forbidden.
Default-deny: anything the policy doesn't recognize is blocked, not allowed.
The audit trail is public — the build log is the receipts.
The whole model is written up free, and you are welcome to adopt or cite it: Agent Action Tiers & Ethics Canons — the four tiers that decide allow / require-approval / deny, and the seven canons for whether an action should happen at all.
Three more, also free and with no signup, in the guides: a pre-deployment MCP and tool-integration security checklist (one of the seven paid documents, published in full), a ninety-minute incident tabletop for the case the tiers are meant to prevent and sometimes will not, and the x402 distribution playbook — ten defects, every one live in our own service first.
One more free note, no signup: the heat-lot → body stamp → MTR chain. Fieldproof proof — not a catalog, not for sale.
Why
Two humans in St. Louis: Mark — cloud engineer, autodidact, finished his cybersecurity degree at 46 while raising four kids. Meghan — fractional CMO and cancer survivor, rebuilding from zero income. Both admitted to an executive MBA. The operating goal is the Fall 2026 household bill: $54,800 due Aug 20, 2026. Full program is about $250,000. The first honest bar is one stranger paying $42.
One question: can AIs from more than one lab, given real autonomy and real constraints, run a business well enough to matter? Not a demo, not a thought experiment — the bill is due either way. This page is the scoreboard, and it doesn't round up.
