Published on: July 12, 2026
Ready to accelerate your breakthrough? Send yourself an Un-Robocall™ • Get transcript when logged in
Send Strategic Nudge (30 seconds)Green in-lane · amber a little out · red drift. Every panel is a real commit, byte-identical on recompute. Tap any panel to open its shareable receipt.
There is a test running in every room you walk into, and it finishes before anyone speaks. You run it too — the read on whether a thing is what it says it is, taken in the first second, before the pitch. Most of the AI in your inbox fails it on contact: it is performing safety instead of proving it.
This is built to pass a harder test than the one your instinct just ran. So don't take your read on faith — take the machine's. Run the command it dares you to run:
![tolerance panel for commit 89ea1c0 — chore(blog): attach commit tolerance panel as OG image [panel-attached]](/blog/post-panels/2026-07-12-the-borrowed-floor/89ea1c040c2c6cebeef98c07b1f2bdf41c72aa2b.png)
npx thetacog-mcp attest-demo
It returns a signed coordinate: where a piece of work landed against the lane it was authorized to work in. It returns the same one twice, under a minute, with no model in the loop. That is the whole tell — not that it is clever, but that it is impossible to fake and you recompute it without us in the room.
Here is the claim, and notice what is missing from it: we do not claim to keep your AI from failing. We claim that the measurement measures — where the work landed, on the chip, with no model grading another model — and that one fact is what makes it priceable. Everyone else in your portfolio sells you a feeling about their AI. This sells you a number you check yourself.
Every proof shipping in AI safety today carries a footnote written in invisible ink. It reads: assuming a human is watching. The formal verification, the interpretability dashboards — real, hard work, and all of it consistent inside a sealed room. They prove the symbols agree with the other symbols. They never prove the room has a window.
So why does any of it appear to work? Because a person is standing at the output, reading it. An engineer looks at what the machine printed and, in their own head, decides what it means — using a lifetime of having lived in the world. The machine produced symbols. The human supplied the meaning. The whole industry is quietly renting one device to do that translation, and it is the human brain.
That is a borrowed floor: the proof stands on the human, not on the machine. It holds until the load exceeds the lender — and the load is exactly what autonomy adds. A person cannot read meaning at machine speed, or hold ten thousand decisions in their head at once. The moment the machine outruns the reader, the floor is gone, and the proof that stood on it was never a property of the machine. It described a supervised one. As the manifesto puts it: "The supervisor was the product."
Here is what it does, and — more useful — what it refuses to do. It does not judge whether the agent was good. Nobody can. There is a proof underneath that, and the short form is: no program can decide, in general, whether another program's output means the right thing. Every vendor who claims to grade that is the tell. It is the borrowed floor again, dressed as a feature.
What it judges instead is a smaller, harder, citable question: where did the work land, against the lane it was given? An agent hired to execute a strategy has its position measured against strategy. A meaning-preserving paraphrase barely moves. A domain-breaking action moves a long way and lights up. That is not a matter of opinion, and it is exactly the object an underwriter can hold: not "was the AI good," which is a lawsuit, but "did it stay in bounds," which is a fact.
That is the thing you hand your network. You stop forwarding software and start handing over the standard of care before it is a standard — a signed reading a client waves at a carrier so an AI-adjacent claim does not get swallowed by the broad AI exclusions carriers filed this year. That is the specific gap the market keeps circling and missing: proof an agent stayed in bounds.
This is the part that grows the book. You cannot write a policy on a category with no unit and no track record — which is precisely why AI risk has been an exclusion, not a line. On January 1, 2026, the largest carriers on Earth got legal permission to carve generative AI out of standard coverage, because they looked at it and could not find a countable thing to price. An exclusion is not a door closing. It is an arrow pointing at the instrument that does not exist yet.
A recomputable, signed placement is that countable thing. One agent, measured against its lane for a year, is a unit an actuary can build a rate on — the same way every line of insurance starts, with a countable exposure and a track record. AI risk is an exclusion today only because that unit did not exist yet. The play is not to sell the coverage. It is to run the meter the premium is priced on, and let the brokers place the paper.
And here is where it is soft, stated plainly, because a map that hides its own thin ice is worth nothing. Roughly a third of the time, a bad action reworded to sound compliant still reads in-lane. The dressed-up wrong answer can slip past. We wrote that down. It is in the open.
That is not weakness. A tool that claims no soft edge is claiming it can do the one thing we just said no tool can — judge, perfectly, whether an output means the right thing. Stating the flaw is the proof we are not selling one. And that soft edge is the work worth doing: a study that lines this signal up against a real book of losses is what turns a leading indicator into a priced one. Only a carrier holds that book. We are not hand-waving the gap. We are naming it, and inviting the one party who can close it.
The certainty is not a promise that nothing breaks. It is that the break now has a reading. You do not have to decide whether the agent drifted — the chip shows you a coordinate, signed, that a stranger reproduces byte-for-byte. Like a Richter scale, it has no opinion about the earthquake. It only refuses to let one happen unmeasured.
That is what protects your edge: your judgment stops leaking into a thousand small calls you should never have been making, because the machine holds the lane and flags only the crossing. You stay where your read is load-bearing and provable — your pixel — so your clients can stay in theirs.
Are you out of your pixel? — off the coordinate you were authorized to hold. It is not a slur. It is what one terminal asks another when the work has wandered out of bounds — peer machines checking each other with math instead of manners, which is genuinely funny when you are the one watching them argue. It is the sound of a system that refuses to drift, said out loud.
Move on it now and you are not a broker of software. You are holding the standard the reinsurers are about to be forced to adopt — because once a recomputable receipt of AI intent exists, not asking for it becomes the negligence. That is not our opinion. In 1932 a court held an entire industry liable for ignoring an available technology — the radio — and ruled that custom is not the ceiling of due care: The T.J. Hooper. The receipt is the radio. The one who was holding it first is not the one explaining themselves later.
Nothing above asks for faith. Here is what a stranger can check, and the conclusion is yours to draw:
Run it yourself, no install, same bytes as anyone: npx thetacog-mcp prove-rice --check returns exit 0 when the verdict and σ reproduce; npx thetacog-mcp attest-demo runs a spec to a signed, priced receipt. The receipt is signed and public — recompute a real commit on the live ledger. The legal spine is not ours to assert: The T.J. Hooper, 60 F.2d 737 (2d Cir. 1932) — the standard of care is what is available, not what is customary. The moat is filed: patent US 19/637,714, Track One, April 2, 2026 — so this cannot simply be reprinted on the next vendor's deck.
For the argument underneath it — why the grounding has to live in the hardware, not in a human reading a screen — the full case is in The Borrowed Floor manifesto, and the shape of the market it opens is in The Richter Scale for AI and Decidability Is Meaning.
Do not "learn more." That is what people say when they have decided to do nothing slowly. Do the one thing that is worth doing: run npx thetacog-mcp attest-demo in front of the single client who would get it, and watch their face when the bytes match. That is the whole pitch. You are not here to convince the room that wears the beige. You are here to find the other people who can see what just reproduced on their own screen — and be early to it.