The Referee Seat Nobody Can Sit In¶
Companies are handing real decisions to AI. The piece nobody has built is an independent way to check whether the AI actually stayed inside the limits it was given. That job, an outside referee who keeps score, is empty on purpose, and it has to cover AI on a screen and machines in the physical world alike.
In one week, three credentialed people each defined the category of keeping AI under control. Real credentials, real arguments. And without meaning to, all three left out the same thing.
- An investor saw a buying problem: pick the security vendor with the most features.
- A security strategist saw a product problem: the tools are scattered, so buy one built for the job.
- A governance officer with her own branded framework saw a design problem: stack the right layers and you are covered.
They are all right, as far as they go. And they are all standing on the same field. Investors picking vendors. Vendors selling coverage. Experts drawing diagrams. Industry groups writing definitions. It's crowded, capable, and well funded.
Here is the thing none of them can do.
None of them can say, independently, whether the AI actually stayed inside the bounds it was given. Not whether the rules exist on paper. Whether they held, in the moment the AI acted, measured by someone who does not also sell you the controls.
That is not a missing feature. It's a structural vacancy, and the chair stays empty for reasons that more money and better products do not fix. Why? Three of them.
- A company cannot grade its own homework. If the firm that sells you the AI also issues the report card on whether the AI behaved, you do not have a measurement. You have a marketing claim. The boxer does not get to referee his own fight.
- An industry group cannot do it either. A trade group is a room full of competitors. A real grade means somebody in that room scores lower than somebody else, so the room quietly agrees to publish definitions instead of grades. That is why the big shared scorecards keep getting announced and never actually arrive. The moment the members have to rank each other, the agreement falls apart.
- A score is a different kind of thing than a framework. A framework is advice. A score is a claim someone can point to, rely on, and be wrong about, with consequences attached. One carries real risk. The other does not. So the industry keeps producing checklists and maturity charts, and stops just short of the one thing that would actually answer the question: did the AI stay in bounds or not.
We have watched this exact story play out before. Banks did not get to set their own borrowers' credit scores, so FICO became its own company. Companies issuing bonds did not get to rate their own bonds, so Moody's did that job. Manufacturers did not get to certify their own products as safe, so UL did. Every time machines start making decisions at scale, the same empty chair appears, and it is never filled by someone already in the game. It gets filled by an outsider with no horse in the race, because being an outsider is the entire point. The independence is not a nice-to-have on top of the score. The independence IS the score.
Now look closer, because there is a second empty chair hiding inside the first.
Everyone I just described is watching software. An AI that approves a payment, edits a file, sends a message. All of it happens on a screen.
But AI already walked off the screen. It is also a drone deciding to hold its position over an airport. A car deciding to change lanes. A sensor system deciding what it is looking at. When one of those acts outside the limits it was given, the cost is not a deleted file you can restore from backup. It is a crash, a near-miss, something physical that you cannot undo. Same question. Same empty referee chair. Far higher stakes.
And here is what makes the chair so hard to fill. The independent referee this moment needs has to keep score on both kinds of AI, the one on the screen and the one out in the world, because more and more they are the same system. Nobody currently in the game can even see the physical side. They were built to watch software. A referee who can only see half the field is not a referee.
So the real question under every conversation about controlling AI right now is not which vendor wins. It is who is allowed to keep score across everything AI now touches, on the screen and off it, and why it cannot be any of the people currently asking you to just trust them.
That question is why I have spent the last eighteen months building an answer instead of a product.
The answer has three parts, and none of them is a sales pitch.
The first is a plain set of rules for what "under control" actually means, written down so anyone can argue with them: a system should not be allowed to grade itself, to be the only witness to what it did, to sign off on its own safety, or to keep running after someone has hit the stop button. And the limits have to be set before the AI acts, not explained after. I call these the Five Laws of AI Governance, and they fit on one page.
The second is a way to turn those rules into an actual score, so "is this AI under control" stops being a sales claim and becomes a number you can check. It's called AQ Score™, filed with the U.S. Patent and Trademark Office as a measurement standard.
The third is the working machinery underneath it: a control system for software AI that is already running and already listed as a reference in a federal standards program, and a sensing system for the physical world that is already running in the field and already feeding those same checks, with the work that joins the two halves filed in the patent record, because the chair that matters has to cover both.
I am not going to tell you I'm the referee. Anyone who declares himself the referee is just a louder vendor, and you have plenty of those already. What I will tell you is that the chair is real, it is empty for reasons that are built into the problem, and the work it takes to earn that chair, an independent score, written rules, machinery that actually runs, covering both screen and world, is work that has to exist before anyone can sit there.
So when something goes wrong, on a screen or in the real world, and the only question that matters is whether the AI stayed inside the limits it was given, ask yourself who you want answering it. The company that sold you the AI. Or someone with nothing to gain from the answer, and the reach to see the whole field.
I know which one I would trust. I built toward it on purpose.
Frequently asked¶
Why can't an AI security vendor measure its own governance?
Because the grade is self-validating. When the company that sells the controls also issues the score on whether those controls held, the market cannot trust the result. Independent measurement requires structural separation from the operator, the same separation that makes FICO independent of lenders and UL independent of manufacturers.
Why don't industry coalitions produce a governance score?
A coalition is a table of competitors. A real score implies winners and losers among its own members, so coalitions retreat to neutral taxonomy and best-practice guidance. That is why shared-body scorecards are repeatedly announced and rarely shipped.
What makes a score different from a governance framework?
A framework is advice. A score is a falsifiable, attributable claim that a third party relied on, with consequences if it was wrong. The liability profile is fundamentally different, which is why most of the field stops at severity ratings and maturity models.
Why does AI agent governance now span both digital and physical domains?
Autonomous agents no longer act only on software. They hold airspace, drive vehicles, and act on sensor data, where an out-of-bounds action causes a physical event that cannot be rolled back. An independent scorekeeper has to measure conformance across both domains, because increasingly they are the same deployment.
What is AQ Score?
AQ Score is a measurement standard for whether an AI system stayed inside the limits it was given. It turns the question "is this AI under control" into a number you can check, scored independently rather than by the company that sells the AI. It is filed with the U.S. Patent and Trademark Office as a measurement standard and covers both software AI and physical autonomous systems.
What are the Five Laws of AI Governance?
The Five Laws are a plain set of rules for what "under control" actually means. A system should not be allowed to grade itself, to be the only witness to what it did, to sign off on its own safety, or to keep running after someone has hit the stop button. And the limits have to be set before the AI acts, not explained after. They are published as a one-page reference.
What is runtime governance?
Runtime governance means the limits on what an AI is allowed to do are enforced in the moment it acts, not reviewed after the fact. Most of what gets called AI governance is process (how an organization adopts AI) or accountability (who is responsible after something happens). Runtime governance is the layer that decides, before the action, whether the AI is allowed to take it.
Who should measure whether an AI agent stayed within its authority?
Someone with nothing to gain from the answer. The company that sold you the AI cannot credibly grade whether its own AI behaved, and an industry group cannot rank its own members. Independent measurement requires a party structurally separate from both the operators and the rule-writers, the same separation that makes FICO independent of lenders and UL independent of manufacturers.