Originally published on Thomas’s Substack. Reproduced from the supplied publication export. Statements and patent-status references reflect the original publication date.
Editorial introduction · added September 17, 2026
Before you read
An engine's test-bench performance does not describe the whole car on the road. This essay uses that distinction to question what AI benchmarks leave out when participants, services, permissions, and routes change together. SSOAR enters as a proposed way to govern that larger interaction. The cost model is an architectural argument, not a measured deployment result.
SSOAR means Session-Scoped Orthogonal Authority and Routing.
Why read it?
- Engineering teams
- Ask which production conditions a model benchmark leaves untested.
- Infrastructure investors
- Separate model efficiency from the operating cost of the system surrounding it.
This introduction is separate from the original essay.
Go to the original essay ↓A Narrative About Why the Most Powerful Machine Ever Built Has Never Left the Garage
For a decade, the engineers and evangelists who built AI systems believed one thing with near-certain conviction:
If it runs on the dyno, it runs.
So they ran it on the dyno.
Over and over. Every benchmark. Every leaderboard. Every demo. Every keynote. Every paper. Every press release.
The numbers were real. The performance was real. The intelligence was real.
Nobody was lying.
The dyno does not lie.
It measures exactly what it measures, under exactly the conditions it creates, with exactly the load it applies, against exactly the resistance it generates.
And the car ran beautifully.
Horsepower through the roof. Torque curves that made engineers weep. Efficiency ratings nobody thought possible five years ago. Token generation so fast it felt like thought.
So they made it faster. And then faster. And then cheaper. And then smaller. And then they put it in everything.
And the keynotes got bigger. And the valuations got larger. And the papers multiplied. And the benchmarks kept falling.
And the car never left the garage.
Because the dyno is not the street.
The street has potholes the dyno never modeled. The street has crosswinds the dyno never generated. The street has other drivers the dyno never simulated. The street has pedestrians who require accommodation. The street has jurisdictions that change at the county line. The street has load that shifts on curves. The street has weather.
The street has construction. The street has a school zone at 8am and a freight corridor at noon.
The street has participants the dyno never counted. The street has authorities the dyno never consulted. The street has modalities the dyno never tested. The street has governance requirements the dyno never measured. The street has transport constraints the dyno never simulated.
And the moment the tires hit asphalt, the numbers change.
Not because the engine lied. The engine did not lie.
But because the dyno measured the engine. The street measures the system.
And the system includes everything the garage left out.
Every participant the AI must coordinate with. Every authority it must satisfy. Every jurisdiction it must respect. Every accessibility requirement it must meet. Every transport boundary it must cross.
Each one independently governed. Each one requiring its own negotiation. Each one multiplying against the others.
Not adding. Multiplying.
That is what the street costs.
And no dyno number tells you that.
Because the dyno runs clean tires on a smooth drum in a temperature- controlled room with a single measured load.
The street runs the same tires on asphalt that has not been repaved since 2009, in crosswind, with a load that shifts every time the speed changes, through six jurisdictions before lunch.
The talking motorheads are not lying either.
They believe what the dyno shows them. And the dyno shows them something real.
But they are describing a car that has never driven.
They are telling you what it will do on roads they have never measured, in conditions they have never modeled, alongside participants they have never counted, under authorities they have never consulted.
They are promising you the street using numbers from the garage.
And the street does not care about the dyno.
The street is the Internet of Everything.
One hundred and fifty billion devices. Participants in every jurisdiction. Modalities in every language. Authorities in every regulatory domain. Transports across every protocol.
All of them present simultaneously. All of them governed independently. All of them waiting at the end of the driveway.
The car is magnificent.
Nobody disputes that.
But magnificent in the garage is a different claim than functional on the street.
No matter what you build in the garage, it is just a two-ton paperweight until it runs on the street.
And the street has been waiting a long time for someone to do that math.
Here is a start.
If the car eats all the gas, you are not going very far no matter how smooth the road is.
When a system must coordinate across independently governed participants, modalities, features, authorities, and transports, the coordination cost scales as a product of those dimensions, not a sum.
Cfrag = k × P × M × F × A × T
Where P=Participants, M=Modalities, F=Features, A=Authorities, and T=Transports.
At the scale of the Internet of Everything, that product does not become a rounding error.
It becomes the dominant consumer of every watt the system touches before a single useful output is produced.
The dyno never measured that.
The street will not forgive it.