Published on: June 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 leading reinsurer's head of AI insurance recently put the industry's position plainly: AI is, at its core, a statistical system, and it carries model uncertainty which splits into reducible and irreducible risk — so insurance can transfer that risk and let enterprises adopt AI at scale. If you run a book, that framing is comforting, because it's the box you already know: build better models to shrink the reducible part, and price the irreducible part off a distribution. It is also, quietly, the most expensive assumption in the market — because the split only works if there is a decidable event to count.
Reducible and irreducible are both species of risk — measurable variance over a known event space. The AI failure that bankrupts a book is not in that space. It is a different animal entirely, and the box doesn't have a wall for it.
There is a century-old distinction the box ignores: Frank Knight's split between risk (you know the distribution, you can price it) and uncertainty (you cannot even enumerate the outcomes). "Reducible and irreducible" lives entirely on the risk side. But the catastrophic agentic failure — capable work executed in the wrong domain, the agent that flawlessly does brain surgery when it was hired for plumbing — is not a draw from a known distribution. It is the agent stepping off the manifold you underwrote, into a space you never specified. That is Knightian uncertainty, and there is a hard reason it cannot be reduced: by Rice's theorem, no system can decide the semantic properties of another system from inside the same failure domain. An eval is software; it shares the agent's failure domain. You cannot model your way out of an undecidable property.
You cannot split uncertainty into "reducible" and "irreducible" when the underlying event is undecidable. There is no distribution to estimate — only a black box you are guessing the tail-dependence of with a more sophisticated guessing machine. Priced confidently, that is the 2008 failure mode wearing an AI costume.
Here is the move. We do not try to decide whether the work is good — that stays undecidable, and it stays the model's job. We decide where it landed: which competence lane it occupies, relative to the lane it was contracted for. We project the intent (the spec) and the execution (the work) onto a fixed 144-node actor-and-patient lattice and read the king-move distance between them. That is a placement of two finite texts on a finite grid — recomputable by anyone, offline, with no model in the loop. "Did the agent stay in its lane" is decidable (the full mechanism and the blind-validation numbers are laid out in The Decidable Slice of Alignment). And the instant a question becomes decidable, leaving the lane becomes a countable event — which means it has a frequency, and a frequency is the one thing an actuary can actually write a line on.
This is how you convert Knightian uncertainty into countable risk: not by modeling the black box, but by drawing a decidable fence around the lane and counting the crossings. You don't price the undecidable. You price the departure, which is decidable.
The crossing is recorded as a signed drift receipt — coordinates, a variance score, and one-way hashes of the ask and the work, never the work product itself. The platform proves the work stayed in (or left) its lane without ever custodying a byte of it, and a counterparty recomputes the receipt offline and trusts no one. Parametric AI covers already settle on measurable performance data; the receipt is that datum — the decidable event your trigger reads. And it is real, not a promise: on a sealed, blind, cross-domain held-out, off-domain work was rejected 10 out of 10, with a 0.90 separation; 0 of 7 adversarial forgeries were accepted; the whole thing recomputes on a laptop.
The honest fence, named not buried: at the finest grain the sensor reads structure more than relation (a role-swap scores 0.76, a paraphrase 0.42), so we certify domain membership — decidable — not within-domain relation, which stays the model's job. And today's frequency is calibrated on our own lived ledger (breach near 13%); the next milestone is blind held-out folds. The fence is the asset.
A primary carrier in Hartford writes the policy; the reinsurer is the capital reservoir that backstops the catastrophic tail. That hierarchy is the whole strategic point. If a carrier adopts a decidable drift receipt, they price one book better. If the reinsurer accepts the decidable slice as the valid way to turn Knightian uncertainty into countable frequency, it becomes the standard of care: every primary carrier downstream is structurally pushed to adopt the same 144-node lattice to get their AI books reinsured at all. The reinsurer doesn't just buy a tool — it sets the ruler the entire market beneath it has to measure by. That is the difference between selling a feature and defining a category.
T.J. Hooper, 1932: once an available device exists, industry custom stops being a defense. The moment a decidable drift receipt exists, "we used a statistical model" is the tugboat sailing without a radio — and the reinsurer is the one who decides when the radio became mandatory.
If you run an AI book, here is what the decidable slice lets you do. Stop pricing the undecidable and start pricing the departure: write the agentic-AI line every carrier is excluding, off a measured crossing frequency instead of a guess. Sell the upside as an option on a competence authority staying in-lane. And — if you are the reinsurer — own the standard of care itself, the toll road every downstream book has to cross. None of it asks you to trust us: the entire claim, the weld, the forge-test, the held-out, recomputes on your own machine in about twenty minutes. Paste the page below into a coding agent and let it tell you, in your own words, whether it holds.
The undecidable part stays the model's job. The decidable part — did it stay in its lane — is now a signed, recomputable, countable event. That is the slice a reinsurer can build a standard on. See it yourself: thetadriven.com/pixel.
