The thesis
One question frames everything on this site. It is asked here, and answered at the end.
On this page
I · The problem
No system in service can.
Not: Has not. Cannot.
The same is true wherever AI now shapes or executes a decision: in a command, a bank, a hospital or a control room. Not the edge. The whole estate: every model, every agent, every actor, human, machine or AI.
What is left is self-attestation: trust, not proof. A system acts, and then it vouches for itself. Nothing weighed the action against authority, and the record is self-made and self-judged. When an inspector general, a coalition partner, an examiner or a court asks what happened, the reviewer is asked to trust the thing being judged.
Look at what the digital world built. The actor certifies itself. The system keeps the record, holds the record, and produces the record on request, in its own format and on its own timetable. Nearly every audit log, compliance report and governance dashboard in service today works this way. We took the one arrangement the Republic was founded to refuse and made it the default of the digital world. Nobody argued for it. Nobody decided it. It was the cheapest thing to build.
While a human being stood in the decision path, the record never carried the weight. The person did. Remove the person and the record must bear a load it was never built for.
A log the actor keeps is not a judiciary. It is a diary.
The pressure
of enterprises will demote or decommission autonomous AI agents by 2027, due to governance gaps identified only after production incidents occur.
A leading industry analyst firm, press release, 26 May 2026agents or more in use at an average global Fortune 500 enterprise by 2028, up from fewer than 15 in 2025.
A leading industry analyst firm, press release, 28 April 2026of organizations think they have the right AI agent governance in place.
A leading industry analyst firm, press release, 28 April 2026The phrase that matters in the first of those findings is identified only after production incidents occur. Detection after the fact is not governance. It is archaeology.
II · Where the answer comes from
The people building the frontier systems have said this year, on the record, that the world lacks the technical and governance tools to pace them. Two hundred and fifty years ago the founders faced the same problem, fought for the answer, and built it into a republic.
A machine that acts and then vouches for itself is the self-attestation of a king. The founders named that defect and built a country to prevent it.
Consider the claim itself. A system that acts, and then issues the only record of having acted, makes a political claim rather than a technical one, and an old one: I am the actor, I am the witness, and my account of myself is the account you get. A free people refused that claim in 1776. Everything that followed from the refusal, the Declaration, the Constitution, the Bill of Rights and the cases resting on them, is the answer they built so that no one would have to accept it again.
First, the ground. Rights are not issued by power. They are held before any government exists and are not power’s to grant — men are “endowed by their Creator with certain unalienable Rights,” and governments are “instituted among Men, deriving their just powers from the consent of the governed.” Declaration of Independence, 1776 Legitimacy comes from outside the thing being legitimized. A power that is the source of its own authority has none.
Then the indictment. When free people set out why they would no longer be governed by a distant will, they did not argue in the abstract. They listed the particulars, and one of them was this: “He has made Judges dependent on his Will alone, for the tenure of their offices, and the amount and payment of their salaries.” Declaration of Independence, 1776 A judge who answers to the party appearing before him is not a judge. That grievance wrote verifier-independence into the founding record, and a war was fought with it on the list. Jefferson’s remedy was a right that does not rest on the word of the party being judged.
Then the remedy, built into the architecture. The Constitution did not ask any officer to be trustworthy. It divided the power so that none had to be. U.S. Constitution, 1787 Madison, arguing for it that year, stated the rule it rests on: “No man is allowed to be a judge in his own cause, because his interest would certainly bias his judgment, and, not improbably, corrupt his integrity.” James Madison, Federalist No. 10, 1787
He also named why the difficulty is structural rather than moral. “In framing a government which is to be administered by men over men, the great difficulty lies in this: you must first enable the government to control the governed; and in the next place oblige it to control itself.” James Madison, Federalist No. 51, 1788 That is the artificial intelligence problem, set out before there were machines to set it for. First enable the system to act. Then oblige it to stay inside what it was permitted. The second half is the hard half. It has never been solved by asking the first half to behave.
Then the guarantees. The Bill of Rights made two of them explicit and both matter here: no warrant issues but upon probable cause, particularly describing what may be searched and seized — a warrant being a magistrate’s prior authorization, never the officer’s own; and powers not delegated are reserved, so what was not granted is not held. U.S. Const. amends. IV and X, 1791 For machines the warrant matters most, because it is review before the act. A machine decision cannot be governed by review after the fact. The act is done, and judgment that comes after it is history, not governance.
Only then the courts, which interpret this inheritance and do not found it. In 1798 Justice Chase named the acts a legislature cannot rightfully commit, among them “a law that makes a man a Judge in his own cause… It is against all reason and justice.” Calder v. Bull, 3 U.S. 386, 388 (1798) In 1948 the Court held that the inferences behind a search must be “drawn by a neutral and detached magistrate instead of being judged by the officer engaged in the often competitive enterprise of ferreting out crime.” Johnson v. United States, 333 U.S. 10, 14 (1948) And in 1995, quoting Madison’s sentence, it called the rule “a mainstay of our system of government.” Gutierrez de Martinez v. Lamagno, 515 U.S. 417 (1995)
That is the order, and the order is the argument. Rights first, because they precede power. The indictment second, because it names what unchecked power does. The architecture third, because a remedy must be built and cannot be promised. The guarantees fourth. The cases last, because a court can only apply what it was handed.
Read it forward and the machine is in it. The King asked to be taken at his word. The answer was a revolution that rejected a king and began the search for a better architecture. A self-certifying system asks for exactly what was refused, in a medium the framers did not have. The claim is not new. Only the claimant is.
So we are not proposing a principle. We are applying the oldest one we have to the newest thing that acts. Stated for this age, the inheritance reads: no power, human or artificial, may be the sole judge of its own conduct.
This is the two hundred and fiftieth year of American independence. The claim the Declaration refused is being rebuilt in silicon — at a scale and a speed the framers had no way to picture, and by default rather than by decision. The answer does not have to be invented. It has to be built, and built at machine speed, because that is the speed the claim now travels at.
None of this is an accusation. Self-attestation is lawful, and it was sufficient for as long as a human being stood in the decision path, because the accountability rested on that person and not on the paperwork. It stops being adequate at the moment the person is taken out, and not before. Lawful and insufficient are different findings. This is the second.
III · The three branches
The American constitutional settlement separated three functions that had been fused in a crown, and gave them to parties with different interests, different tenures and different sources of authority. The rule is written by one, the action is taken by another, and the judgment as to whether the action was within the rule belongs to a third that can do neither.
What carries across to a machine decision is not the imagery but the structure: the same three functions, held apart for the same reason, applied to a new actor. The parallel is not exact and need not be. It holds where it counts, on who may judge.
Congress writes the rule and appropriates the means.
An appropriation is a grant with limits, and spending past it is unlawful on its face. The machine equivalent is the authority set: delegated authority, commander's intent, rules of engagement, credit policy, the standard of care. Written before the case arises, and written by somebody other than the party who will act.
The branch built for speed, initiative and action.
Hamilton wanted energy in the executive and designed for it. An AI agent is that energy — capability and tempo, acting under real authority delegated to it. What it does not hold, and what the executive does not hold either, is the authority to rule on whether its own act stayed inside the grant. It proposes, and once permitted it acts. It does not judge.
Applies the rule written in advance to the case at hand.
Independent of the party before it. Applies the authority another party wrote, not a judgment of its own; decides permit or deny against that authority, and says so in a written instrument that outlives the moment and can be examined by somebody who was not there. It cannot propose and it cannot act, which is the reason it can be trusted to judge.
A judiciary fails the moment the party before it can unseat the judge who rules against that party, diminish the judge’s pay or overrule the judgment. That is the whole of constitutional design, not a fine point of it. The failure is as complete when the court’s records are kept in the defendant’s basement and produced at the defendant’s pleasure.
That is what verifier-independence means. It is the test most systems fail, not because the engineering is hard but because almost no one set out to be checked by a stranger. The evidence is held by the party whose conduct is in question. Review requires that party’s cooperation. The reviewer is asked to take the actor’s word.
The Constitution enumerates what the federal government may do and reserves everything else. The architecture carries a default, and the default is no. A power that was not delegated is not held, and the burden falls on the party claiming it to point to the grant.
A machine should hold authority the same way. Not a general license bounded by a list of prohibitions, but a specific grant, and the ability to show which grant a given action was taken under. Deny by default is not caution. It is the constitutional position.
The third branch has no agenda. It cannot propose and it cannot act. It does not decide what should be done, it has no view on the mission, no ambition of its own, and nothing to gain from the outcome. It decides one question only: whether what was proposed is inside what was granted.
That restraint is why it can be trusted with the role, and it is the commercial point as well. An adjudicator does not compete with the systems it governs, and it has no reason to slow the mission, because it is not trying to do the mission. The agent-capability scales in use measure how much a system can do on its own. This sits off those scales by design. The further the industry travels along them, the more it needs something that is not on them.
“[The judiciary] may truly be said to have neither FORCE nor WILL, but merely judgment.”Alexander Hamilton, Federalist No. 78, 1788
The architecture occupies the role of a third branch. ProofSmith builds it and holds no authority under it: ProofSmith never decides. The authority is the customer’s, and the Arbiter decides only what that authority already allows. The architecture supplies the separation.
We sit between the brain and the trigger. The proposer proposes. The Arbiter decides. The claim stops there. The Arbiter is neither judge of the case nor jury of the facts. It rules on one thing, whether the act was inside the authority you wrote. It does not make a model more accurate, it does not judge whether the authority behind a decision was sound, and it does not judge whether the facts before it were true. Those remain human judgments.
The false choice
“Enterprises are treating AI agent governance as binary, either locked down or fully trusted, and that is the root cause of failure.”An analyst at a leading industry analyst firm · press release, 26 May 2026
That is the choice an organization faces when the only instruments available are permission granted in advance and a log kept afterward. Restrict the system and lose the capability. Trust the system and lose the accountability. Organizations tend to oscillate between the two and call it governance.
A constitution takes a third position. Do not ask whether the actor can be trusted. Do not forbid the act. Adjudicate it: at the moment it is proposed, against authority written in advance, by something with no stake in the outcome and nothing of its own to advance.
IV · Two tests
ProofSmith is the trust layer for digital information. It establishes provenance, decides what may access or transform an artifact — permit or deny, on authority — and emits independently verifiable proof of that decision, whether the actor is a human, an AI, or a machine.
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.
The evidence is on The Solution: one hundred and seven systems assessed against these two tests, and what we found.
V · The posture
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.
Assists inside a task a person is still running.
Runs whole workflows alone, inside limits a person set.
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.
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 their 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.
What we take from the study and our own count of one hundred and seven systems against the two tests are on The Solution.
And the instinct that a control slows you down has it backward. The pattern is older than us, and it holds in every arena where going forward carries risk: the safeguard that looks like a limit is what licenses the advance.
The flight envelope is what lets the aircraft be flown hard. The wingman covers your six.
Posture is a wrestler’s word. On your heels you are vulnerable. Balanced and leaning forward, you are set to attack, and the balance is what makes the attack safe to launch.
The lesson of Ford v Ferrari: if you want to go faster, give me brakes. Brakes are what let you carry speed into the corner.
Audited books built the capital markets, debt and equity alike. A risk that can be evaluated can be priced, and a risk that can be priced can be funded. That is permission to lever up and grow, and the foundation of wealth creation.
Enforceable contracts made strangers safe to deal with. A promise that can be enforced can be relied on, and a record that can be admitted is what makes it enforceable. That is what let commerce reach past the people who knew each other, and grew the economies that outlasted their rivals.
Provable authority shifts command from back on its heels to forward-leaning: delegate at speed, and own every call.
One principle, every arena. The safeguard is never the brake on aggression — it is the license for it. ProofSmith brings that safeguard to the decision itself: any proposer proposes, the Arbiter decides on the authority you wrote and signs it, so command, or the board, can lean forward at speed without surrendering accountability.
When the question comes — from an inspector general, a coalition partner, Congress, or from an examiner, an auditor, a court, a shareholder — you do not reconstruct an answer. You stand in front of them with one, signed. That is your six.
VI · The close
We opened on one question: who governs the governor in the age of the machine?
No commander should answer for a decision they cannot prove.
No nation should field machine intelligence it cannot hold to account.
We govern the decision, and we prove it — independently verifiable, accountable to the authority behind it.
Not merely declared — shown.
The demonstration
We would rather show it than describe it, and build it around your problem rather than our example. Tell us which of the six is costing you, and that is the one we will build.
Start a conversationSee the six problemsDownload the two-page summary
We are not asking you to trust us. We are asking you to let us prove it.
In detail
Someone acted. Nobody can prove the action stayed inside the authority that was granted. The grant lives in one system, the action in another, and the two were never joined at the moment that mattered.
An AI contributed to an action and there is no record of what it proposed, who accepted the proposal, or on whose authority it went forward. Accountability for the outcome stays with the organization regardless.
The record of the decision is held by the same system that made the decision. Reviewing it requires that party's cooperation, in that party's format, on that party's timetable. The reviewer is asked to take the actor at its word.
Where the question gets asked
Authority, intent and rules of engagement are written with great care, and then the record of whether they were honored is kept by the system that acted.
Read →Grid, water, pipeline, rail and plant operators are automating control decisions faster than they can evidence them.
Read →Credit, trading, payments, onboarding and surveillance decisions are moving to models and agents. The obligation to evidence them did not move anywhere.
Read →Clinical decision support, prior authorization, triage and documentation are automating. Liability and the standard of care are not.
Read →Conformity declarations, management assertions and authorizations to operate all rest, in the end, on records the signer produced about the signer.
Read →