Published on: July 10, 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.
Before the argument, the artifact. Run this and read what comes back:
npx thetacog-mcp attest-demo
It returns a signed, recomputable coordinate — where a piece of work landed on a fixed 144-cell lattice, the σ, the lane, the on-chip signature — the same receipt this post carries, byte-identical every time you or anyone else runs it. Nothing here asks you to take that on faith; it is the first thing you can check.
Here is the claim, in its full attackable form, and notice what is missing from it: we do not claim we can stop your AI from failing. We claim only that the measurement measures — semantically, on-chip, with no model in the loop — and that single fact is what makes it fair, priceable, and decidable. Every AI-safety vendor promises the thing we refuse to promise: that their layer will prevent the bad output. That promise is unverifiable — Rice's theorem says no program decides a non-trivial semantic property of an arbitrary program — and the moment you make it, you have handed the buyer's security officer the edge case that kills the deal. We deleted the promise. What is left is a ruler, and a ruler is the only thing the people with the money were ever actually waiting for.
This is not a thought experiment; it is a calendar. On January 1, 2026, a set of ISO commercial general liability endorsements went live — CG 40 47, CG 40 48, and CG 35 08 — that carve generative AI out of standard liability coverage entirely. State regulators approved more than 80% of the AI-exclusion filings from subsidiaries of Berkshire Hathaway, Chubb, Travelers, and AIG. Read that plainly: the largest carriers in the world looked at autonomous-agent risk, decided it was unpriceable, and got legal permission to walk away from it. The coverage your competitors assume they have is being deleted at renewal.
The second clock is the regulator's. On August 2, 2026, the EU AI Act's obligations for high-risk systems bind — and high-risk explicitly includes creditworthiness assessment and insurance pricing — on top of DORA, in full force since January 2025, which already demands that a bank prove operational resilience rather than assert it. Both regimes want the same thing and neither will accept a policy document as proof: evidence of control, not a promise of good behavior.
Two gaps, one shape. The carrier cannot price a subjective failure. The regulator cannot accept a subjective proof. Both are starving for the identical missing object: a countable, objective reading of where an agent actually is. Capital is already rushing the coverage gap — a new specialty category, hundreds of millions in fresh funding — but everyone racing in is selling coverage into the hole. Nobody is selling the ruler that makes the hole measurable. That is the asphalt everyone drove past.
The tempting pitch — the one a founder reaches for because it sounds strong — is prevention: our system detects the drift and halts the agent before it acts. Do not make it. It is the single sentence that converts a friendly technical review into an autopsy. The buyer's CISO does not have to disprove your architecture; they only have to name one input where the agent slips the leash, invoke Rice, and the deal is dead by lunch — because you promised a guarantee that no software auditing software can mathematically hold.
So we removed it. Read our claim for what it does not say: it does not say the agent will be stopped, corrected, aligned, or made safe. It says the agent's position is measured — a coordinate on a fixed lattice, produced with no LLM anywhere in the computation. That LLM-free-ness is not a performance detail; it is the whole game. A model grading a model is subjective, un-auditable, and re-drifts for the same reason its target does. A gzip-grounded compression distance against a fixed geometry is none of those things: it is a deterministic function of the bytes, so two strangers on two machines get the same number, so an actuary can price against it, so a regulator can accept it. Fair, because no model's opinion is in the loop. Priceable, because it is a count and not a grade. Decidable, because placement — unlike correctness — is a thing a program can actually decide.
The cleanest way to say what we do is the way a surgeon says it. Here is the operating room. Here are its walls — the declared lane the agent is supposed to work inside. We do not hold the scalpel and we do not stop the surgeon's hand. We do one thing: if the work leaves the room, the telemetry registers the departure — objectively, on-chip, in a form a stranger recomputes. A human then has to check. And if a human did not check when the instrument said the boundary was crossed, that is the moment liability attaches — not to us, to whoever ignored the reading.
That is a ruler, so hold it to a ruler's standard: an error bar you can see. On a sealed, blind, cross-domain corpus the sensor separates in-lane from out-of-lane at 0.90 and rejects out-of-domain ten out of ten. Against a scrambled null — the same bytes with their meaning destroyed — the signal stands at 4.48σ. And we publish the inch we have not closed, because hiding it would be the tell: paraphrase-invariance sits at 0.30, a surface reword still moves the reading more than we would like. We say so out loud. A Richter scale with a stated tolerance is an instrument. One that claims perfection is a fraud. Everything past this line — a checker that auto-halts, a human-in-the-loop fallback, a router that saves tokens by short-circuiting on a clean read — is something you build on top of the signal. It is not ours to promise, and we do not.
Picture the Head of Commercial Liability Underwriting at a top carrier. She spent eighteen months lobbying regulators for the exact exclusion that just went live, because she could not underwrite a hallucination — a qualitative event that takes years of litigation to even define. She opens our spec expecting one more alignment wrapper, and here is the turn in her own head:
"Every vendor tells me they can stop AI drift. I know they are lying — you cannot guarantee a halt on an unpredictable execution, which is precisely why I wrote the exclusion. But this one is not claiming a kill-switch. It is claiming a decidable placement receipt. In-lane, out-of-lane, signed, recomputable, no model in the loop. That changes the object I am pricing. I do not have to insure 'the AI behaves.' I can write an affirmative carve-back where an in-lane failure is a covered, bounded event and a lane-jump is a countable alarm — the thing I actually refuse to eat. The premium I have Fortune 500 clients begging to pay, at nearly the risk I can now bound. I need my actuaries on this telemetry today, before Berkshire does."
She does not adopt it because it is clever. She adopts it because it hands her back a market she was forced to abandon — the exact coverage the exclusion created demand for — and it does so by turning the unpriceable half into a boundary she can draw.
Now the reinsurance chief underwriting officer, the one the capital markets keep calling about a half-billion-dollar AI catastrophe bond. She has told them, every time, that it cannot be built — because a parametric bond needs a trigger like the Richter scale needs a magnitude: an objective number that decides the payout with no lawyer in the room. There has never been one for AI. Who rules that the model failed — a judge? You cannot securitize a lawsuit. She opens the technical spec:
"A parametric trigger is the one thing I have never had for this asset, because 'the AI messed up' is a subjective claim and subjective claims are un-securitizable. But this is a coordinate on a fixed 144-cell lattice — deterministic, git-show-identical, signed on-chip, and it does not ask one model to grade another. It is a mathematically recomputable baseline. That is my magnitude scale. I do not need it to stop the bad trade; I need it to prove, immutably, that the boundary was crossed. If the coordinate reads out of bounds, that is a countable event — the bond pays on the reading, not on a claims adjuster's opinion. I can structure on this. I need the licensing rights to the signal locked before a competitor does."
The bond does not care whether the agent is good. It cares whether the trigger is objective. That is the entire reason prevention was always the wrong pitch and measurement was always the product: capital does not buy safety, it buys a countable event.
The third desk is a broker's general counsel, trying to sell AI-liability policies to clients terrified of a class action. His problem is older than AI: how do you defend against a negligence claim when there is no agreed standard of care to point to? He finds the patent, the live feed of hundreds of real receipts, and the standard of care assembles itself in front of him:
"Today, if a client's agent causes harm, they get sued for negligence — and there is no settled standard for what 'careful' even means with an autonomous agent, so the defense is a coin flip. But this is the radio case. In 1932, T.J. Hooper: two tugs sank because they had no weather radios, radios were available and affordable, and the court held the whole industry negligent for not adopting them even though almost no one had yet. The mere existence of a working, affordable instrument set the standard. So here is my deposition, a year from now: plaintiff's attorney holds up this repo and asks my client, 'Were you aware a decidable, cryptographically signed measurement of agent drift existed, and cost almost nothing? Why did you deploy without it?' My client has no answer. Ignoring the ruler is the negligence. I do not sell this as software. I mandate it as a condition of coverage and bake the fee into the premium — it is the only thing that survives that question."
He is right about the mechanism, and it is the quiet engine under the whole thing: a measurement does not need adoption to create duty. It needs only to exist and be reachable. The scale's existence is what manufactures the standard of care — which is why these desks move now, not after the market catches up.
There is a reason a Richter scale is not a proprietary product: a magnitude scale owned by one company that no one else can verify is worth nothing to the bond market. Ubiquity is the precondition for financialization — Wall Street will not build a half-billion-dollar instrument on a black box a single startup controls. Which resolves the strategy cleanly, and it is the opposite of hoarding. The ruler wants to be everywhere. So the value does not sit in the measurement; it sits in the toll booth on the instruments built from it.
The split is clean along the line of who is doing what. If you are an operator — filling contracts, building a checker, routing a map-of-maps where every intent gets its own 144-lattice — the signal is yours to use and improve, freely; we are not going to chase the people doing the actual work for pennies. If you are a financial issuer — pricing, triggering, or underwriting a bond, an option, or an insurance policy on the signal — that is the toll booth, and it carries a commercial license on the instrument you build. The operators get ubiquity and zero friction. The issuers get a standardized, recomputable data feed they could never have built themselves. There is more than enough in options, bonds, and insurance on competence that we never need to bill the builder — it is already baked into the premium downstream. Hand the ruler to the public and you do not lose the moat; you manufacture it, because a standard of care that everyone can reach is exactly the one everyone is now negligent to ignore.
A toll booth needs a unit, and the unit is what turns a fee into a market. One execution of the lattice is one intent — one discrete, verifiable piece of cognitive work — and a fixed bundle of them is a standard quantity: call it an Agent Year, the kilowatt-hour of cognitive labor. A power company does not ask how many watts your dinner will take; it reads the meter. An underwriter does not need to know how many intents an operation will burn; they price the Agent Years the meter counts. And because an agent runs at some multiple of human speed, the unit is decoupled from the calendar entirely — a server cluster can turn over a decade of cognitive labor in a weekend, and the meter does not care about the clock, only the count. That is what the instruments are denominated in: the operator burns Agent Years for free, the issuer who writes a policy or a bond against them pays the toll on the unit, and the aggregate of live receipts becomes a drift index the market prices speed against accuracy on. The precise rate is a licensing-page term, not a blog claim — but the shape is a kilowatt-hour, and whoever owns the definition of the unit owns the meter.
None of the boardroom economics happens unless the meter is everywhere, and meters do not spread because a CFO mandates them — they spread because the person at the terminal wants them. So meet the person who makes the whole thing ubiquitous: the engineer running six agents in Cursor or Claude Code, blocked in staging for three weeks because the compliance team wants an LLM-as-a-judge evaluation pipeline bolted onto every action. That pipeline adds hundreds of milliseconds of latency, costs a fortune in tokens, and — this is the part that grates — it is still software grading software, so it proves nothing. Here is the turn in that engineer's head:
"I have been stuck because risk wants me to prove my agent will not drift, and the only tool anyone offers is another model watching my model. Then this: npx a harness that runs in CI and measures the execution against a fixed 144-cell lattice — no round-trip to a judge model, 0.90 separation, a decidable pass or fail. It is a compiler check, not an eval. I do not write fifty brittle edge-case tests; I wrap the agent in the harness, point Cursor at it, let the AI code until the tests pass, and hand compliance a signed placement receipt on every commit. It is open source — free to run locally and in CI. The license only bites when Finance wraps the signal in a bond, which means the standard is paid for on the back end, not out of my budget. No SaaS approval, no alignment argument. I merge to main."
That is the Rosetta Stone: the three roles that never agree in a modern enterprise all say yes to the same artifact for their own reasons. The developer wants velocity and gets an npx harness that lets an AI coding assistant verify its own work. The risk officer wants a defensible standard of care and gets a recomputable ruler. The underwriter wants a transferable unit and gets the Agent Year. The developer adopts it because it is the path of least resistance to shipping; because they adopt it, the telemetry is embedded in the enterprise; because it is embedded, the broker finally has something to underwrite. Velocity at the terminal is what funds the market in the boardroom — the same signal, read three ways.
The fence is the product, so read it as carefully as any promise. We do not claim to stop your AI from failing — Rice says no one can, and we do not pretend otherwise. We do not claim to certify what your agent's output means or whether it is correct — meaning is the subjective, litigable thing we deliberately refuse to sell, because the instant we sold it we would be back inside the undecidability we escaped. We claim only placement: where the work landed relative to its declared lane, decidably and recomputably. And we publish the limit of even that: paraphrase-invariance at 0.30 is real, and it means a determined surface reword can still move the reading — an open inch, stated, not buried.
This restraint is not modesty; it is what makes the thing bankable. An instrument that promises exactly what it can keep, and names the edge where it cannot, is one an actuary can put in a distribution and a regulator can accept as evidence. An instrument that promises the undecidable is one that gets caught the first time a confident, in-lane, wrong decision sails straight through it. We would rather hand you a ruler that is honestly 0.90 than a kill-switch that is dishonestly 1.0 — because only one of those is something a stranger will insure.
Nothing above rests on our say-so. The AI-exclusion wave is on the record: ISO forms CG 40 47, CG 40 48, and CG 35 08 took effect January 1, 2026, and state regulators approved a majority of the AI-exclusion filings from the largest carriers — the coverage gap in Section B is a filed fact, not a framing. The regulatory clock is public: the EU AI Act's high-risk obligations bind August 2, 2026, and DORA has demanded demonstrable operational resilience since January 2025. Rice's theorem — no program decides a non-trivial semantic property of an arbitrary program — is the result behind Sections A, C, and I; the halting problem is its familiar special case. Normalized compression distance (Cilibrasi and Vitányi, Clustering by Compression) is why the reading is a count and not a grade. T.J. Hooper (1932) is the negligence-by-availability precedent behind Section G. And the measurement itself is filed under US Patent Application 19/637,714. The economics past this post live in the receipt is the fiduciary duty and you cannot clone a coordinate.
Three things to actually do, in order of commitment. (1) Run npx thetacog-mcp attest-demo, then browse the live receipts — each one ringing exactly where the work landed, green in lane, amber bleed, red drift — and hold the reading to your own null; that is the whole proof, and it costs a minute. (2) Read the open standard (Protocol 144) and the agent-year to see the unit the meter counts, then decide whether placement, not correctness is the decidable slice your risk model has been missing. (3) When you would rather price against a ruler than argue about a verdict, license the signal per agent — the operator's use is free; the toll booth is only for the instruments you build on top.
The strongest thing we ever did for this product was subtract from it. We stopped promising to control the AI, because that promise was a lie physics would not let us keep and a wall the buyer's own engineers would happily push us into. What remained was smaller and unkillable: the measurement measures, on-chip, with no model in the loop, and the boundary is drawn where a stranger can read it. That is the Richter scale. It does not stop the earthquake. It just makes the earthquake a number — and a number is the one thing a carrier, a reinsurer, and a court have all been unable to get and unwilling to move without.
Two clocks are running — coverage deleted at renewal, the regulator's deadline weeks out — and both are starving for the same missing object. This is not pressure; it is arithmetic. The receipt is on the table, it costs a stranger one second to turn over, and the moment it exists, not checking it is the thing read aloud in the deposition. Your pixel is the coordinate your work is supposed to occupy — the lane you declared and the carrier priced. So the only question the ruler ever asks is the one a stranger can now settle in a second: are you in your pixel, or are you out of it?
npx thetacog-mcp attest-demo. Every load-bearing claim in this post traces to something you can run or a source you can check — the exclusion forms, the deadline, the theorem, the receipt. We handed you the ruler. What you do when it reads out of bounds is the only part that was ever up to you.