Published on: June 26, 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.
We believe you are leaving the largest untapped premium pool of the decade on the table, and you are doing it for one reason: you have been told agentic-AI risk is unpriceable, so you are waiting for a standard to arrive before you write the line. That wait is the mistake. The line is writeable today. Not after a working group publishes a framework, not after a regulator blesses a methodology, not after three years of loss data accrue — today, off first principles that already exist and a breach number that is already being lived. This is not a thought piece asking you to imagine a future market. It is a market-making call: the door is open right now, and the first carrier through it prices the peril everyone else is still excluding. Whoever writes the line first owns the tollbooth before the rest catch up. The peril, named plainly, is an agent wandering out of the lane it was hired for — the question underneath every one of these claims is simply: are you out of your pixel?
The reason agentic AI has been "unpriceable" is not that the risk is exotic. It is that no instrument existed to count the loss event. The instrument exists now. The moment a countable event exists, a frequency exists; the moment a frequency exists, a price exists — and a priced peril is a line you write instead of exclude. The only question left is who writes it first.
Walk the chain you already run, and you will see exactly where it broke for this peril. A premium is a frequency times a load. A frequency is a count of a loss event over an exposure base. A count requires a countable event — something you can point at and say, unambiguously, "this one happened, that one did not." And a countable event requires a decidable detector: a rule that returns yes or no on each case without a human re-litigating it every time. That chain is not philosophy; it is the operational floor under every line you have ever priced. Auto has the accident report. Property has the claim. Mortality has the death certificate. Each is a decidable detector that produces the count that produces the frequency that produces the price.
Agentic AI had no such detector, so the chain never started. There was no rule that returned a clean yes or no on "the agent did the thing we are insuring against," so there was no countable event, so there was no frequency, so the peril was excluded. Read that sequence carefully, because the conclusion is the opposite of the one the industry drew. The peril was not excluded because it is uninsurable. It was excluded because no instrument existed to count it. Those are different diagnoses with opposite consequences: the first says stay out forever, the second says the day an instrument arrives, the line opens. The instrument is the thing that was missing — and the instrument is what changed.
You did not exclude agentic AI because you judged it uninsurable. You excluded it because the detector that produces the count did not exist, and you cannot price a frequency you cannot compute. That is a tooling gap, not a property of the risk. Close the tooling gap and the exclusion has no reason to stand.
Here is why no one could build the detector, and why that failure was real rather than lazy. Rice's theorem (1953) is unforgiving: no program can decide a non-trivial semantic property of another program in general. "Is this agent good, is it safe, is its reasoning sound" — every one of those is a semantic property, so every one is undecidable. That is not a gap a bigger model or more compute closes; it is a wall. If the only question on the table is the semantic whole, the detector is impossible, and the exclusion is forced. So far the pessimists are correct, and you should respect that they are.
But the catastrophic question is narrower than the semantic whole, and the narrower question sits below Rice's line. We do not ask whether the agent is good. We ask whether it stayed in the lane it was hired for — domain placement, not quality. That question is finite: a fixed text projected onto a fixed 144-node lattice, where the answer is the king-move distance (Chebyshev — the largest single-axis leap) between the lane it was hired for and the lane it actually worked in. A finite text on a fixed lattice with a computable distance is a decidable property by construction. It is sub-Turing; Rice never reaches it. It runs in about fourteen milliseconds, recomputes on a laptop, and — this is the part that matters to your trust in the number — no model is in the loop. The detector is not another opaque AI judging the first one. It is a deterministic geometric measurement anyone can re-run and get the same answer. A decidable property has a decidable detector, and we built the detector for the one question that bankrupts the insured.
Rice forbids deciding whether the agent is good. It does not forbid deciding where the agent worked. "Did it stay in its lane" is domain placement — a finite, deterministic, model-free measurement below the undecidability line. That is the decidable detector the whole chain was missing, and it is the only one the price actually needs.
Let me name what you actually had on your hands before the detector, in your own vocabulary, because this is the concession that reframes everything. Frank Knight's distinction is the cleanest tool in the cabinet: risk is uncertainty with a measurable probability distribution; uncertainty is the kind you cannot measure, where no distribution can be fit at all. And I want to be precise and fair, because the lazy version of this argument insults you and you would be right to reject it. You are not a naive summer of probabilities. You have copulas, tail-dependence coefficients, Expected Shortfall — a sophisticated apparatus for joint extremes and fat tails. None of that is the issue. The issue is more specific and it is upstream of all of it: every one of those tools is a probabilistic model of continuous variables calibrated on history, and the agentic peril is discrete, ahistorical, and — until the detector — undetectable. You cannot fit Expected Shortfall to a loss you cannot count. You cannot calibrate a copula on a history that does not exist. Your apparatus was never wrong; it was starved of the one input it requires — a countable event — and so this peril was never risk in your sense at all. It was Knightian uncertainty wearing a number, which is exactly why it felt unpriceable and exactly why the exclusion felt forced.
The decidable ruler is the thing that converts it back. The moment "did the agent leave its lane" becomes a decidable yes-or-no, the event becomes countable; the moment it is countable, you have a frequency; the moment you have a frequency, you have a measurable distribution, and Knightian uncertainty collapses back into Knightian risk — the priceable kind. Nothing about your apparatus changes. The copulas, the tail-dependence, the Expected Shortfall all still apply, now fed a real frequency instead of a guess. We do not replace your math. We hand it the input it was missing.
Before the detector, agentic risk was not risk — it was uncertainty wearing a number, undetectable and therefore uncountable, and no amount of copula sophistication can price a loss you cannot count. The decidable detector makes the event countable, which makes the distribution real, which converts the peril from Knightian uncertainty back into priceable risk. Your apparatus does not change. It finally gets fed.
A frequency you forecast is a story; a frequency you lived is a book. So here is the one you can build on. We run the in-lane detector on our own agents, all day, every commit — not as a demo but as the way the work actually ships. That ledger is not a projection; it is observed loss experience. The in-lane breach rate — how often the agent did capable work in the wrong lane — lands near ten percent, with a ninety-five percent confidence interval of roughly seven to sixteen percent, across about a hundred and fifty-nine signed attestations. That is a real frequency on a real exposure base. It is a loss ratio waiting for a load. The premium prices the frequency and loads for the semantic volatility of the work being insured — and, crucially, it never loads for measurement precision, because the measurement is deterministic and free. You are not paying for our confidence in the ruler. The ruler is exact; only the world it measures is volatile.
And the proof you price off never asks you to trust us. Each attestation is an ed25519-signed bundle of coordinates and one-way hashes — where the agent's intent landed, where its delivery landed, the king-move between them — and nothing else. The work product itself never leaves the insured's machine. We call this oracle, not host: the carrier prices the peril without ever custodying a byte of the insured's intellectual property, and anyone can recompute the receipt offline and get the same coordinates. Change one field and the signature breaks. There is no third party to trust and no IP to leak — the number you price off is one you can check yourself, not a fear you take on faith.
This is not a forecast you have to believe — it is a lived loss ratio you can recompute. Breach near ten percent, signed and observed on a real ledger, with the proof carried as coordinates and hashes so the insured's IP never leaves their machine. You price the frequency, load for volatility, and never pay for measurement noise because there is none.
Now the part that turns "you can write this" into "you should write this first." There is a doctrine in liability law you already know: The T.J. Hooper (1932). Two tugs lost their barges in a storm because they had no radios to hear the weather warnings. The industry custom was not to carry radios — yet the court held the tugs negligent anyway, because an effective safeguard existed and a reasonable operator would have used it. Custom is not a defense once a better safeguard is available. Adoption of the safeguard does not wait for the industry to make it standard; the availability of the safeguard is what makes its absence negligent.
Apply that here and the whole motion inverts. Once a decidable in-lane detector exists — once it is possible to know whether an agent stayed in its lane and to prove it with a signed receipt — not measuring becomes the negligent act, even before the industry adopts it as standard. That means adoption is not a sales motion you have to push; it is a liability mandate that pulls. The enterprise that deploys agents and cannot show it measured them is the T.J. Hooper tug without a radio. And for you, the carrier, the consequence is sharp: the first underwriter to write the line off this standard is not just early to a market — it is defining the standard of care that every later policy, claim, and court will be measured against. Exclusion is stable only while no one can price. The instant one carrier prices the slice, every excluded policy is premium a competitor will take, and the standard belongs to whoever moved first. That is the tollbooth. It is built once, by the first mover, and the rest pay to pass.
Once an effective safeguard exists, failing to use it is negligence — even before the industry adopts it (The T.J. Hooper, 1932). So a decidable in-lane detector makes measurement a liability mandate, not a sales motion. The first carrier to write the line does not just enter the market; it authors the standard of care everyone else is later judged by. That is the tollbooth, and it goes to whoever moves first.
Three honest questions, because a claim this size has to answer them before you move a dollar. Why is there no alternative? Because the only other move inside the behavioral frame is to keep excluding, and exclusion is provably not protection — Rice forbids deciding the semantic whole, so the sole remaining path to a price is to carve out the decidable where and measure that. There is no third option that produces a count. Why is this not too good to be true? Because it is bounded and it says so out loud. It prices domain placement, not quality — the undecidable whole stays excluded, exactly where it belongs. The separation is measured, not asserted: cross-domain separation of 0.90, ten-of-ten blind off-domain rejection, a scrambled-null floor around 4.48 sigma, a role-swap-versus-paraphrase gap (0.76 against 0.42) proving the detector reads domain and not surface wording, and bounded-autonomy holding two of two. And the honest edges are named, not buried: the breach number is calibrated on our own ledger today, so out-of-sample experience will sharpen it; and in-universe surface-robustness — paraphrase invariance — is the open inch we are still closing. A measurement names its error bars. A story does not. Why has nobody else done it? Because the entire field has been looking in the wrong place — inside the black box, trying to certify the agent's reasoning, which is the undecidable problem Rice already closed. The move here is almost embarrassingly orthogonal: stop looking inside, place a rigid geometry around the output, and the question becomes decidable. It is not a smarter version of interpretability; it is a different problem with a different answer.
So the test is the opposite of "trust us." You do not have to believe the ten percent, the geometry, or the doctrine. You run the detector on your own agents, watch the live king-move between documented intent and delivered reality, and recompute the signed receipt offline yourself — reliability and practicality are not claims you take on faith, they are the two things you verify in the first ten minutes. If you want the deeper theory afterward — why the exclusion is itself the liability, and how the same primitive becomes a tradeable option — it is laid out in The Exclusion Is the Liability and Black-Scholes Didn't Touch the Stock. The only thing you need to do first is stop waiting for a standard that has already arrived: find the pixel your agents are supposed to land in, and measure whether they stayed there. The market is open today. Walk through the door before the carrier across the street does.
Run it yourself, do not take our word for it: the detector, the live drift, and the recomputable receipt are at thetadriven.com/pixel. The line was never uninsurable. It was only ever uncounted — and now it counts.
