Published on: July 25, 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.
A broker said the most useful sentence anyone has said to us this year: "I'm a bit confused on what you are offering here." She was right, and the confusion was ours. We had been serving her a tool to resell when her job is to write a class — and we had handed her a caveat that misread our own product, telling her that whether staying in-lane correlates to fewer claims was a validation study we still wanted to run. That sentence gave her a reason to wait. It was also wrong: parametric cover has never required a proven correlation to loss. Hurricane policies pay on measured wind speed. So here is the meal re-plated, with the caveat removed and the trigger where it always belonged.
Every course below is plated the same way, because the plating is the argument. First the maître d' presents the dish — pure flourish, tongue slightly in cheek, the course named in the market's own terms before a word of argument reaches the table. Then the inner monologue: the exact sentence the course is built to make you think, written down before the course is served. That is not a wish about your reaction; it is a prediction you get to grade. Then the mechanic that forces the sentence — the specific reason you can't simply decline to think it. Then the ingredients. Publishing the prediction in advance is trust-inversion at the scale of one section, and it is the opposite of manipulation: manipulation needs the dark, and this is printed on the menu before you taste anything. If a course ends and its sentence didn't fire in your head, the course failed and you caught it — and catching it is the meal working anyway.
The win condition, declared before the first plate: not your agreement. This meal wins if you leave the table and recompute — run the command, check whether the second run matches the first, take the trigger language to someone who binds. It fails if you leave merely nodding. Grade us on that.
The maître d', presenting: Recorder, Rewound Twice — served sealed, playback lever on your side of the pass. The kitchen could rewind it for you; it declines to. What it replaces is the post-incident report assembled eleven months later from application logs that can say the agent emitted 40,000 tokens and cannot say what it was trying to do.
Inner monologue it should trigger: "They handed me the playback and dared me to check it myself — so nothing after this needs their word."
The mechanic — why it can't be ignored: the command runs on your machine, where we cannot reach, and it takes about ninety seconds. A dare, once made, is irreversible — run it and the verdict is yours, decline it and you now know you declined. Either way you cannot return to the moment before the lever was offered.
Don't take a sentence of what follows on faith. Run npx thetacog-mcp attest-demo on your own machine right now, and here is exactly what comes back: a signed placement verdict — in-domain, out of domain, or unplaced — for a piece of work, computed with no model anywhere in its path. Then run it a second time on the same input and compare. That second run is the whole product, and it is why we get to say "why we believe" without asking you to believe anything.
The ingredients on this small plate, and they only work cold and together: the playback lever, which costs you ninety seconds and proves we are not hiding the ball; the identical second run, which is a property rather than a promise; the halo of letting you catch us in the lie; and the empty chair — the argument we very much have, held visibly in reserve until course I. What you should feel is not "these people are impressive." It is "wait, I'm the one holding the verdict."
A flight recorder is not valuable because it is clever. It is valuable because the investigator who reads it does not have to trust the airline.
The maître d', presenting: Telemetry Tartare — served raw and entirely surface: hard-braking counts, swerve events, minutes after midnight. Season it however you like; it will never tell you why the driver turned. What collapses without the mechanic is the incident review where every dashboard was green and the loss still happened.
Inner monologue it should trigger: "Every AI monitor I've been shown is doing to outputs exactly what telematics does to driving — counting symptoms and staying blind to intent."
The mechanic — why it can't be ignored: the exclusions are already being drafted. That is a date in the past, accruing, and reading this does not pause it — every quarter your market spends unable to price the exposure is a quarter it writes the exposure out of the policy instead.
Telematics never made anyone a better driver by understanding drivers. It measured braking, swerving, and speed — surface symptoms, recorded after the fact, structurally blind to what the driver intended. That was enough to price a motor book, because a car has a small number of ways to fail and they all leave physical marks.
Now look at what the AI observability market sells: keyword filters on generated text, token counts, latency, API call frequency, and a scan of the output after the model has already produced it. That is the same category error, one level up. It is non-semantic telemetry — it measures the exhaust and infers the engine. And unlike a car, an agent has an unbounded number of ways to be wrong, most of which leave no mark in a token count at all.
This is why your market is writing aggressive AI exclusions rather than aggressive AI policies. You cannot underwrite a tail risk from symptoms. An underwriter asked to price a black box, monitored from the outside, after the fact, correctly declines. The exclusions are not timidity. They are the honest response to an unpriceable object.
The failure isn't that AI is risky. It's that every instrument pointed at it measures behavior rather than meaning — and behavior, post-hoc, is the one thing an underwriter cannot turn into a number.
The maître d', presenting: Exclusion Clause, Chilled — plated on your own letterhead, served at the temperature of a renewal call. The dish it replaces is the submission that came back with AI carved out of it and a client asking you what changed.
Inner monologue it should trigger: "I have already had this exact conversation — I just had it about cyber, ten years earlier."
The mechanic — why it can't be ignored: this is checkable against your own book this week. Open the last renewal where an AI-touching exposure got narrowed, and read who asked for the carve-out. It was not the client.
You have lived this arc once. Early cyber was uninsurable for exactly this reason: nobody could measure the thing being insured, so the market priced fear, then wrote exclusions, then — slowly — built the measurements that made a real class possible. Firewalls, logging standards, incident response retainers, the whole apparatus. The class did not appear because cyber got safer. It appeared because cyber became measurable at a boundary an underwriter could name.
AI is at the pre-measurement stage of that same arc, and the tell is identical: the exclusion arrives before the product. When a market cannot price something, it does not stay silent — it writes the thing out and waits. Your clients experience that as their AI project stalling in legal, and they blame procurement. It is not procurement. It is that no instrument exists to bound what they are asking you to place.
The connection is not that we have a product for you. It is that you have seen exactly how this ends, and the missing piece both times is the same shape: a measurement at a boundary, not a judgment about behavior.
The maître d', presenting: Parametric Pour, Measured at the Rim — filled to a line etched in the glass, not to the sommelier's judgment of your thirst. What it replaces is the eighteen-month claims adjustment where both sides hire experts to argue about whether the model "should have known."
Inner monologue it should trigger: "A measured boundary is a trigger — I don't need a correlation study, I need the line etched in the glass."
The mechanic — why it can't be ignored: forwarding this costs you nothing. Rhetoric stakes your credibility when you pass it on; a measured trigger stakes nothing, because whoever receives it can check the measurement themselves and their verdict does not reflect on you.
Here is the correction we owe the market, and it is a correction to something we ourselves said badly. We once wrote that whether staying in-lane correlates to fewer claims is a validation study we want to run. That was true as a research statement and completely wrong as a product statement, and it handed a broker a reason to wait.
Parametric cover does not require a proven correlation to loss. A hurricane parametric policy pays on measured wind speed at a named station. It does not pay because wind speed predicts your roof damage; it pays because the parties agreed in advance that the measurement IS the event. The measurement is the trigger by definition, not by prediction. That is the entire elegance of the instrument, and it is why parametric settles in days rather than quarters.
So the contribution is not a monitoring tool. It is a measurable object that can sit in trigger language: a semantic boundary, defined per deployment, attested at execution, recomputable by the carrier's own analyst without us in the room. What you can do with that — and we cannot — is know what a carrier needs to see written before it binds. That sentence is worth more than anything on this page.
The instrument already exists and has for decades. What was missing was a measurement that could stand where wind speed stands.
The maître d', presenting: Second Signal, Plated Beside the First — two dishes arriving together, because neither means anything alone. The pairing replaces the quality score: a number between zero and one that has never once survived a deposition.
Inner monologue it should trigger: "I don't need anyone to measure quality — I need to know when the spec was honored and the result was still bad."
The mechanic — why it can't be ignored: once you have seen the answerable question, the unanswerable one becomes conspicuous. "Was the output good?" cannot be answered by any instrument. "Did it stay inside its declared boundary, and did the operator report a problem anyway?" are two facts, both decidable, and you will notice the difference in every vendor deck from now on.
Out of lane is the ordinary claim. The agent left its declared boundary, the receipt says so, the trigger fires, that case is simple and it is where everyone's imagination stops.
The interesting case is in-lane with a bad result. The spec was respected. The boundary was never crossed. The receipt is clean — and the work is still poor. That combination is not a defect in the measurement. It is information you cannot get any other way: it means the spec is wrong, or the math behind it is wrong, or something upstream changed and nobody noticed. The agent did exactly what it was told, and what it was told was insufficient.
And here is the property that makes it underwritable rather than merely interesting: it cannot be sandbagged. A party trying to manufacture a payout has to produce a boundary breach, which is visible and attributable. Producing the other signal — staying rigorously inside a declared lane while the output quietly degrades — is not something you can fake in the direction of your own benefit. The two signals fail in opposite directions, which is exactly what an actuary wants from a pair of inputs.
Which means we never have to measure quality. We supply one decidable fact — the boundary. The deployer supplies the other — a reported problem, or a business metric they already watch. Their intersection is the event.
The reason this survives Rice's theorem is that it never tries to defeat it. We do not claim to know whether the output was good. We know where it landed, and someone else tells us they are unhappy. The pair is decidable even though neither half is a judgment about quality.
The maître d', presenting: Sandbag, Served Empty — presented open so you can see there is nothing in it. The dish it replaces is the moral-hazard objection raised in minute forty of a meeting that had eleven minutes left.
Inner monologue it should trigger: "Testing this myself is cheaper than constructing the argument against it."
The mechanic — why it can't be ignored: the test costs ninety seconds and one command. Building a rigorous objection to a measurement you have not run costs an afternoon and still leaves you unsure. The asymmetry is the point.
Here is what we do not know, said before you have to ask. We do not know the base rate. How often a deployed agent leaves its declared boundary, across industries, at scale, is not something anyone has measured yet — because until the measurement existed there was nothing to count. That is a genuine gap and it affects pricing, not triggering. A parametric instrument can be written before the base rate is known; it cannot be priced well until it is. Anyone who tells you otherwise is selling.
Moral hazard is the second open question, and it is the one your underwriters will raise first. If a payout follows a boundary breach, does the deployer get careless about boundaries? Our answer is structural rather than reassuring: the boundary is declared by the deployer and attested against their own declaration, so loosening it to farm payouts means declaring a lane so wide that the policy prices accordingly. The gaming move is visible in the artifact that prices the policy. That is an argument, not a proof, and it deserves your scrutiny more than our confidence.
The sandbagging chapter of the book works through why a system that can quietly underperform inside its own tolerance is the hardest failure to catch — the actuarial blindspot is the same problem this course is standing on, one abstraction up.
The maître d', presenting: Receipt, Recomputed Rim to Rim — the same dish made twice, in two kitchens, indistinguishable to the crumb. It replaces the vendor attestation that cannot be reproduced without the vendor.
Inner monologue it should trigger: "If it computes identically on my analyst's machine, its truthfulness stops depending on anyone's good faith — including mine."
The mechanic — why it can't be ignored: reproducibility is a property, not a claim, and properties do not care whether you were persuaded. Run it twice. If the second run differs, this entire post is worthless and you found out in three minutes.
The receipt is generated with no model anywhere in its path. That constraint is not a preference, it is the requirement — a probabilistic component in the measurement would make the measurement unpriceable for the same reason the agent is unpriceable. Same input, same artifact, every time, on any machine.
That property is what lets a measured boundary function as a trigger. A carrier's analyst can take the same commit and the same declaration, recompute the placement, and get the same verdict without our cooperation — which means our cooperation is not a dependency in the claims process. Adversary-recomputable is the standard. Anything less is a vendor asking to be trusted at exactly the moment nobody should be.
And the falsification is offered before the claim: if the two runs ever differ, the instrument is void. We would rather you discover that in ninety seconds than discover it in a dispute.
The maître d', presenting: Fidelity Bond for a Non-Human Hire — one carving knife, one table, one first signature. What it replaces is the enterprise absorbing one hundred percent of an agent's failure because no instrument existed to share it.
Inner monologue it should trigger: "When we hire a person there is a whole social contract behind them. An AI agent has none — and someone is about to write the first one."
The mechanic — why it can't be ignored: there is one carving knife per table. The first broker to put defensible trigger language in front of a carrier defines the shape everyone else's submissions get compared to. This is a race that has already started, and the starting gun was the first exclusion.
When a company hires a person, an entire apparatus arrives with them that nobody thinks about: fidelity bonds, employment practices liability, professional indemnity, legal recourse, and a body of precedent about what happens when a worker causes harm. The risk of hiring a human is bounded, priced, and shared. That apparatus is invisible precisely because it works.
An AI agent arrives with none of it. When it fails, the enterprise absorbs the entire operational consequence, then discovers the loss sits in an uninsurable gap. That is not a technology problem. It is the absence of a social contract — and it is exactly where cars were before motor insurance existed. Cars did not become socially acceptable because they got safe. They became acceptable because society built an instrument that made the risk shareable, and then decided collectively it was worth it.
And here is what makes this version better than a cheque. The natural payout for a semantic breach is not cash in ninety days — it is triage: labor deployed immediately onto the coordinate where the anomaly fired. Engineers, review, attention, funded the moment the trigger fires rather than after adjustment. A business that depends on agents the way it depends on people gets, at the moment of failure, wind in its sails instead of a hole in the hull. The instrument stops being purely a risk transfer and starts being an employment instrument: it pays in the thing that actually fixes the problem.
The question is not whether AI agents will get a social contract. Humans have had one for a century and nobody argues about it anymore. The question is who writes the first trigger language — and that person will be a broker, not a vendor.
The maître d', presenting: Trigger Language, Flambéed at the Table — lit with the match we have been holding since the first course. What it replaces is the industry custom defense, which has already failed in court once and will fail the same way again.
Inner monologue it should trigger: "Whether I like this is irrelevant — if a measurement exists and my industry doesn't adopt it, that's the negligence finding."
The mechanic — why it can't be ignored: precedence outranks preference. In The T.J. Hooper, an entire industry's universal practice was held to be no defense, because the available precaution was cheap and the custom was simply behind. The judo flip is that once one carrier writes a boundary-attested class, not requiring it becomes the exposure — for everyone who didn't.
Here is the authority we have been holding in reserve, and it is not ours — it is a court's. In 1932 a tug lost its tows in a storm. The tug had no radio. No tug in the industry carried radios. Judge Learned Hand held that universal custom was no defense: the precaution was available and cheap, the industry had simply lagged, and "there are precautions so imperative that even their universal disregard will not excuse their omission." An entire industry doing the same thing has never been a shield.
Apply that to a measurement that exists, runs in ninety seconds on commodity hardware, and produces an adversary-recomputable record. Once such a thing is demonstrably available, "nobody in our market required it" stops being a defense and starts being the finding. The book works through this inversion in detail — the T.J. Hooper inversion — because it is the hinge the whole liability argument turns on.
So the flip. Today, adopting a boundary-attested class looks like the risky, early move and waiting looks prudent. That is backwards. The moment one carrier writes it, every carrier that didn't is holding a book they cannot defend as current practice — and every broker who placed without it is in the file. The standard does not need everyone. It needs one, and then it becomes the thing the rest have to answer.
The maître d', presenting: The Bill, Paid in Labor — itemized, and the recipe attached. The wager is not that you agree; it is that you cook it.
Inner monologue it should trigger: "Checking this is cheaper than carrying the question out of the room."
The mechanic — why it can't be ignored: an open loop is expensive to carry and cheap to close. One command, ninety seconds, and the question stops following you around.
Here is the raw material, handed over as ingredients rather than conclusions. Parametric insurance is not new — hurricane, earthquake, and crop parametric products have paid on measured indices for decades, and the entire design principle is that the measurement replaces the adjuster. The T.J. Hooper, 60 F.2d 737 (2d Cir. 1932), is the case on custom versus available precaution, and it is short enough to read in full over a coffee. Rice's theorem is why nobody can promise you an AI behaves well: every non-trivial semantic property of a program is undecidable, which is a mathematical result rather than an engineering complaint. And the EU AI Act and SOX-style separation-of-duties both share one principle worth stealing for this argument: the party being defended cannot certify its own innocence.
Two places to go deeper, both of which contradict us in useful ways if you read them adversarially: the book chapter on the actuarial blindspot, which is about the failure mode this whole instrument is built to catch, and our sibling post The Decidable-on-Silicon Claim We Checked, which contains an explicit list of what we do not claim. Read the not-claimed list first. It is the part that makes everything else defensible.
The to-do is the same one the meal opened with: run npx thetacog-mcp attest-demo, then run it a second time and compare the two. If they differ, we are wrong and you spent ninety seconds finding out. If they don't, you are holding the object this entire argument rests on, and the next question is yours to answer rather than ours: what would a carrier need to see in the trigger language before it binds?
Now grade the meal. Ten courses were served, each with a sentence printed before you tasted it. Count how many actually fired in your head — you are the only person who can compute that number, which is exactly why the win condition was recomputation and never agreement.
