73 assessed
Enforces, no evidence
Category Definition
Why do none of them offer proof?
Two questions decide whether a governance claim is worth anything. Of one hundred and fifty-one systems assessed against them, none passes both. ProofSmith is built to pass both.
Two tests
The two tests, and what we found.
Enforcement. Does the system stop the action, or does it only observe it? A control that records what happened is a witness. A control that determines whether it may happen is a governor. They are not the same instrument and they are not substitutes.
Verifier-independence. Can a third party check the decision without the cooperation of the system that produced it? If the answer is no, the record is an assertion by an interested party, however well signed and however well stored.
For eight months we have searched the public record and put each system we found that claims to govern machine decisions against these two questions. Both tests are defined on this site and the sources are public; the register of placements is ours and is not published. One counterexample settles it. We have not yet found a system that answers yes to both.
Enforces, no evidence · 73 of 151 systems placed here
They enforce. The record is their own.
Identity providers, privileged-access managers, AI gateways and agent-security platforms, together with the industry reference architectures now being published for securing agents. They establish who an agent is and what it may touch, and they can stop it. The record of each decision is kept by the party that made it.
Neither · 38 of 151 systems placed here
They govern. They neither stop nor prove.
Governance, posture, risk and observability tools. They describe, inventory and score, which every program needs, but they neither stop an action nor produce evidence a third party can check.
Evidence, no enforcement · 26 of 151 systems placed here
They attest. They do not decide.
Attestation, provenance and signing tools. They sign a workload, a model, a build or a statement, but they do not decide whether an action may run, and most depend on trusting the issuer or the hardware.
No verifier-independent evidence
Verifier-independent evidence
Enforces the action
ProofSmith
Enforces, with evidence
Does not enforce
38 assessed
Neither
26 assessed
Evidence, no enforcement
One hundred and fifty-one systems assessed, one hundred and forty of them placed on these axes. All one hundred and forty sit outside the upper right; three sit on the line between two of those cells, none of them touching the upper right. A further eleven were adjudicated and set off the axes as belonging to a different layer. If you know of one that belongs in the upper right, tell us and we will place it. ProofSmith’s own place there is a claim until a party you authorize has checked one of our decisions without our help.
Every system outside the upper right is a prospective partner, not a competitor: ProofSmith sits after it, at the last gate before the action runs.
The ProofSmith architecture is patent pending.
Source: ProofSmith adjudication register, version 5.3; publicly available information (product documentation, patents and filings), as of October 26, 2026. The method is on Evidence.
Download the Category Definition (PDF, one page)Download Why ProofSmith (PDF, three pages)
A second count
What a third party measured
A leading industry analyst firm, in a 2026 study of one hundred and seven agentic-AI deployments across nine industries, scored each on six capabilities at five levels, 0 through 4. Level 4 is the top: the system acts with no person in the path, and the firm calls it the level most likely to deliver transformative return. Forty-one deployments were scored at Level 2, sixty-six at Level 3, and none at Level 4.
Level 2 · Basic
41
Assists inside a task a person is still running.
Level 3 · Intermediate
66
Runs whole workflows alone, inside limits a person set.
Level 4 · Advanced
0
Acts with no person in the path.
Deployments analyzed, by level. Source: a leading industry analyst firm, 2026, which notes the examples were not extensively validated.
What we take from it. Level 4 is where the value is, and the return is only half of it. An organization that can delegate authority to a machine and still answer for the result moves at a speed its adversary cannot match, on decisions it cannot otherwise safely automate at all.
Why it stops there. Levels 0 through 3 keep a person in the decision path. That person is the accountability. Level 4 removes the person. Nothing takes his place. The models are capable; the Level 3 deployments already plan, replan and act on production systems. The gap is not capability. No one had built what lets an organization take the last step and still answer for it.
Our own count, of a different set. Of the one hundred and forty systems we placed on the two tests above, none reached the upper right, enforcement with evidence a stranger can check. Two studies, two methods, and in each an empty place at the top: the industry has built capability up to the last step, and has not built what the last step requires.
ProofSmith is not on that scale. The scale measures how much a system can do on its own; an adjudicator does none of it, and the further a deployment climbs, the more it needs one.
10 October 2026
The problem has been named. No platform can solve it for itself.
On 10 October 2026 Satya Nadella wrote that every AI model should be treated as an insider risk, with the controls kept outside the model. He set out seven principles. Here is where we stand on each.
| The principle | Where we stand | Whose layer |
|---|---|---|
| Model diversity. No model checks its own work | Any proposer: human, machine or AI. The proposer proposes; the Arbiter decides. | Ours |
| Observe everything. A tamper-proof record of every action | Every decision, permit or deny, leaves proof as a byproduct of the decision. | Ours |
| Verifiability. Test the whole system, failures included | We test the layer that decides. The customer tests the model. | Shared |
| Independent controls. The organization sets access and actions | Authority is written in advance. The Arbiter permits or denies on it, before the act. | Ours |
| Independent auditability. Validation independent of the AI | Anyone authorized can check the proof without taking the word of the operator, the platform or ProofSmith. | Ours |
| Containment. An authorized person can stop it | When in doubt, the Arbiter denies. The platform holds the brake. | The platform’s |
| Incident disclosure. Tell those affected when it fails | Our records are the evidence a disclosure rests on. The customer makes the disclosure. | The customer’s |
Source: Satya Nadella, “Models as Insider Risks in the Super Intelligence Era,” article on X, 10 October 2026. Paraphrased.
No platform can solve it for itself: evidence that is self-attested is not proof. None of the seven says who, outside the organization, can check its record.
ProofSmith has built the architecture, and has the independence, that solves the problem. A working Arbiter, built and demonstrated, not promised. Proof anyone authorized can check without taking our word. The rule binds us too.
9 October 2026
Who governs the governor in the age of the machine?
On 9 October 2026 David Friedberg, on the All-In Podcast, named the fight that is coming: who gets to be in charge of the AI. Thirty-five minutes on how to make a model obey left that question open.
15:37 to 25:17
Three rules for a model to follow: one line, a terms-of-service clause, Asimov’s three laws. Each is a rule the model applies to itself.
28:56
A bank’s model that stops approving mortgages on its own judgment, like a utility that turns the power off.
1:23:33
An agent told to save money cancels subscriptions and changes a phone plan.
Each leaves the power to act and the power to judge the act in the same hands. One rule offered, do what the user wants unless it is illegal, is the right rule in the wrong place. The proposer proposes. The Arbiter decides.
Source: All-In Podcast (Jason Calacanis, Chamath Palihapitiya, David Sacks and David Friedberg), “Is Claude Conscious? Pope Rejects, Model Welfare Movement, OpenAI’s Math Backlash, France Riots,” episode published 9 October 2026 — youtube.com/watch?v=TBLHdXABwAg. Times are from the episode as published. Paraphrased; no host is quoted.
Where the idea comes from, and our reasoning, is on The Question.
First principles
Self-attestation is not proof.
An agent may act alone. It must never be its own judge and witness.
Enforces, with evidence
ProofSmith decides before the act. The proof needs no one’s word.
The decision is taken before the act, against authority written in advance, and the proof can be checked by anyone authorized without relying on the word of the operator, the platform or ProofSmith.
Start a conversationSee how it works
We are not asking you to trust us. We are asking you to let us prove it.