On September 12 Dario Amodei published We Must Pace the Frontier. Step one of his plan is the one I want to talk about, because I have been running a small version of it for a year without knowing it had a name: evaluators embedded inside the lab, with employee-level access, so that nothing ships without someone who was in the room and could see what the system saw.
That is a governance idea at frontier scale. It is also an implementable schema at any scale. Here is the existence proof from a house.
The system
ML Systems is a Rhode Island construction company with a software layer. Seven AI agents work a house record — the Master Ledger — alongside one human, me. The minds propose roof heights, material grades, values, sequences. The question the whole design answers is: who is allowed to say what is true?
Four rules do the work. I wrote them up before this weekend; they map onto Amodei's step one almost line for line.
1. Claims, not facts. The ledger never stores "the roof is 8.5 ft." It stores a claim: who made it, the evidence grade behind it (measured 90 · sensed 85 · stated 80 · record 70 · modeled 60), and the value. A claim carries the grade of its evidence, not the rank of its author: an inference enters as modeled, the lowest grade; a vision read as sensed; a tape measure as measured. Nothing promotes itself. A grade rises only when a seat adds its own claim beside it.
2. Seats, not permissions. Six parties can author a claim today: custodian · homeowner · vera · cda · pi · record. Five seats can ground one: the homeowner, the town record, VERA (verification), PI (the orchestrator, the one homeowner-facing mind), and the Custodian. CDA — the design swarm — holds no seat, by design. Four more minds are proposed as claimants without seats: a mind may claim; it may never ground. The seats sit inside every output the minds produce, with the same access the minds have — every field, every evidence line — which is exactly the "employee-level access" clause. They are not reviewing a summary after the fact; they are inside the row. Lit means input, not agreement.
3. Domain-scoped authority. Reconciliation is by domain. The homeowner leads on the interior. The record leads on the legal card. VERA leads on the exterior and on anything derived. The Custodian leads every domain. A seat's authority is scoped to where its evidence is strongest, which is what keeps one evaluator from becoming a rubber stamp for the whole record.
4. Content-bound, lapsing signatures. A stamp is two keys — the homeowner's and the Custodian's — bound to a content hash. Change the content and both lapse; the row goes back to unstamped. Stamp or override: an override enters the record as the Custodian's own claim, not a silent edit. And when credible sources disagree, the row's state is conflict; it is quarantined and may not be consumed downstream. Disagreement is a state the schema can hold, not a number it averages.
Why this is the same idea
Amodei's evaluators are there so the people with the power to stop a release are also the people who can see everything the release is built from. The seats do that for a house: the five parties who can ground a claim are embedded in every output the minds produce, with full access, and the record cannot move without them. Pace is not a policy you ask a fast system to honor. It is a decision in the grammar about who may speak.
The scale is different. The property is not.
Labels
No ML Systems deconstruction has been performed yet. The recovery target is modeled; the crane sequence is aspirational. What is measured is the software — a shipped app, a working ledger, and an ontology that seven agents and one human speak fluently.
Further reading
- The ledger design: Claims, Not Facts: Building an Auditable Multi-Author Record for a House
- The ontology: Seven Families, One Ledger: How a House Gets a Grammar
- The founding argument, written this weekend: Pace the Frontier, Pace the House
Top comments (3)
The "claims, not facts" rule is the part that maps least obviously onto Amodei's step one and most usefully onto real systems. Storing the evidence grade alongside the claim, and refusing to promote a grade without a new claim from another seat, makes the audit trail non-forgeable by construction rather than by policy — that's a much stronger property than an approval workflow.
The conflict state is the other good decision; most schemas either average disagreement away or crash on it. One question: when two seats hold equal authority in a domain and disagree, does the row stay quarantined indefinitely, or is there a defined resolution path — tiebreak seat, time limit, escalation? That's the part I always find underspecified when I try to write this for something smaller than a house.
Good question, and the answer is that the design refuses the premise: there are no two seats with equal authority in a domain. Each of the four domains has a lead (the homeowner on the interior, the record on the legal card, VERA on the exterior and on derived values), and the Custodian leads every domain. So the terminal path is always the same person: stamp or override, and an override enters the record as the Custodian's own claim, signed, not a silent edit.
On the timer: nothing in the schema expires a conflict on its own. A quarantined row stays unconsumable until a stamp lands, and it sits in a queue a human can see. That is deliberate. A stale quarantine costs a visible backlog; an automatic tiebreak costs an invisible wrong number.
For something smaller than a house I would keep exactly two things: one named lead per domain, and a conflict state that downstream code cannot read through.
Some comments may only be visible to logged-in visitors. Sign in to view all comments.