Published on: June 29, 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 AI risk looks unpriceable only because the wrong thing was being measured. Whether an agent did its job well is genuinely undecidable, and no honest premium can be written on it — that part the skeptics get right. But there is a second number under every output: how far it drifted from the lane it was hired for, and that number is decidable, recomputable on a laptop in about fourteen milliseconds, with no model in the loop. This post is the consequence of having it. The same number prices AI as an insurance premium for a carrier who wants a floor, and as a variance swap for a desk that wants to hedge. One ruler, two markets — and for the first time the insurance rail and the options rail read off the identical, signed, stranger-recomputable measurement.
The whole post in one line: we never price the quality (undecidable). We price the king-move distance from the contracted lane (decidable) — and that one quantity is simultaneously a breach frequency (the premium) and a realized variance (the swap).
If you carry the risk, you already feel the hole. The underwriter today covers AI with a silent-AI exclusion because the alternative is reserving against a number nobody can compute — and an exclusion is a deferral, not coverage. The chief risk officer signs off on an agent fleet with no signal for whether any of them is still doing the job it was authorized to do. And the loss that actually bankrupts a book is not bad work; it is capable work in the wrong place — an agent flawlessly rewriting your auth when you asked for a CSS fix, answering a tax question as if it were licensed to. It is invisible to every quality check, because the work looks fine. What is missing is not a better quality score. It is a decidable, countable event: did it leave its lane. The instant that is countable, it has a frequency — and a frequency is the one thing an actuary can write a line on.
Here is what the measurement hands you, depending on which seat you are in. If you are the underwriter, it hands you the first AI-competence line you can actually quote — a premium on the rate at which agents leave their lane, with a real number in the reserve triangle instead of a placeholder. If you are a risk desk, it hands you a hedge with a settleable underlying — you can be long or short the turbulence of a lane you do not even own. If you are the AI builder, it hands you a system you can insure and a receipt your customers recompute for themselves. None of this asks anyone to trust a vendor's score; each is something the holder can re-derive. That is the contribution: not a claim about AI, but an instrument other people can pick up and act on.
The capability is a ruler, and the choice inside it is load-bearing. We project the intent (the spec) and the execution (the work) onto a fixed 144-node actor-and-patient lattice, walk the connectivity with a deterministic recursive walk, and then, block by block, take the Chebyshev — king-move — distance (picture a chess king: one step in any direction, so the score is the single worst axis, never a flattering average) to the nearest in-lane reference: distance zero is green (in lane), a step out is amber, too far is red (foreign domain). The aggregate fraction of out-of-lane mass is one scalar, driftPct; a lane's calibrated tolerance band is the strike K (live: 4.668%), and driftPct above K is the loss event. King-move matters because it weighs a diagonal shift exactly like an orthogonal one — the agent must be in-lane on every axis, so "ninety percent compliant in domain, zero percent in role" reads as the breach it is, not a comfortable average. The whole computation runs in about fourteen milliseconds, recomputes on a laptop, and has no model in the loop — a geometric measurement anyone re-runs to the same answer.
This is why it is Chebyshev and not Euclidean or a copula: there is no smooth joint density between "retrieving a document" and "practicing law." There is a wall, and the agent is on the wrong side of it. What kills you is the worst axis, not the blend — so the ruler measures the worst axis.
Once driftPct is a number, the uncertainty in it splits into the two quantities a market actually prices. The first is frequency: how often does a lane breach (driftPct above K)? That is the insurance question — the rate of loss events, the thing a premium is written on. The second is turbulence: how much does driftPct itself swing, window to window? That is the options question — the realized variance, the thing a swap is written on. They are two reads of the same series: the carrier cares how often you cross the fence, the desk cares how violently you move near it. And the live variable sitting over both is the calendar — after August 2, EU AI Act Article 14 requires the signal that a system is still in its authorized role, which turns "nice to measure" into "must produce." The uncertainty is real; the point is that it is now a measured uncertainty, not a vibe.
A premium or a swap is only as trustworthy as the loss history under it, so here is the anchor that makes both safe to price. The number is deterministic: npx thetacog-mcp premium returns the same value every run. The history is tamper-evident: every priced row sits under an ed25519-signed Merkle root, and editing, inserting, or hiding a single breach row changes the root — npx thetacog-mcp ledger-attest --verify rejects it, so the producer cannot quietly Goodhart its own ledger. And the per-job receipts are coordinates and one-way hashes only — the work product never leaves your machine, so a carrier prices the peril without ever custodying a byte of your intellectual property. This is what releases the payment: not a dashboard you trust, but a number you regenerate off a history that cannot have been faked. (Honest fence: tamper-evident, not tamper-proof — the host holds the key; append-time co-signing and on-chain anchoring are the next rungs.)
Step back and see what you become once the measurement exists. The insurance rail reads driftPct as a breach frequency and sells a carrier a floor; the options rail reads the same driftPct as a realized variance and sells a desk a hedge — two projections of one king-move number, the way price and volatility are two reads of one underlying. That is what makes it a market and not a feature: a floor needs a hedge to clear, and a hedge needs a settleable underlying to exist, and both come free from the one decidable quantity. The reader who internalizes this stops asking "can AI be insured" and starts asking "what is my lane's fair vol, and am I long or short it." You move from the person waiting for a standard to the person quoting one.
Why the options rail can exist at all: the underlying is stranger-recomputable. Two counterparties settle a variance swap only if both independently compute the realized variance and agree. An LLM confidence score can never be that underlying — no two parties recompute it to the same value. A king-move on a fixed lattice can.
Here are the live figures, every one recomputable by you, off the same sealed ledger. Rail one, the insurance premium (npx thetacog-mcp premium), over 200 attestations, 182 priceable: breach rate 15.4% (driftPct above K), carried with a Wilson 95% confidence interval of 10.9% to 21.3% — a credible interval, not a false point estimate; semantic volatility (the standard deviation of driftPct — the Black-Scholes σ) 1.87; premium base × p̂_upper × (1 + λ·σ/K) landing at 256 units, priced off frequency and loaded for volatility, and PRICED rather than advisory because the interval is tight enough to write into a policy. Rail two, the variance swap (npx thetacog-mcp variance), over 178 clean attestations, window 20: fair variance 3.234 (fair vol 1.798), vol-of-vol 2.129, latest realized variance 5.489 — above fair, the lane heating up — quoting bid 1.104 / ask 5.363 in variance points, a long marked at 2255 units. And the fence, stated plainly: we do not price whether the work was good — that is undecidable (Rice's theorem), and anyone selling insurance on an AI's execution is selling a solved halting problem. A surgeon who declines to operate is still practicing medicine; one who picks up a wrench and starts on the plumbing is not. We prove the surgeon never left the operating room, and price how reliably they stay in it. Both rails live entirely on the decidable side of that wall.
You do not have to trust a single number on this page — that is the whole design, so go verify it. Run the detector on your own agents and watch the king-move between documented intent and delivered reality; recompute the premium with npx thetacog-mcp premium, the swap with npx thetacog-mcp variance, and the seal with npx thetacog-mcp ledger-attest --verify. The deeper theory is in The Decidable Slice of Alignment, The Exclusion Is the Liability, and Black-Scholes Didn't Touch the Stock. The forcing function is already on the calendar; the instrument is already recomputable. Stop waiting for a standard that has arrived: find the pixel your agents are supposed to land in, and measure whether they stayed there. The market is open today. Are you out of your pixel?
