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.
We believe the alignment problem has been mis-framed as a single, monolithic, mostly-undecidable thing — "make the AI good / make it want what we want" — and that this framing has quietly hidden a decidable sub-problem that you can measure, sign, share, and trade today. The dangerous failure is not bad work. It is capable work in the wrong domain: an agent that competently builds weapons when you asked for crayons, that flawlessly rewrites your auth when you asked it to fix the CSS, that does brain surgery when it was hired for plumbing. That failure is invisible to every quality check — because the work itself is good. Only the coordinate reveals it is off-target. And the coordinate, it turns out, is decidable.
We do not grade whether the work is good (undecidable, and the model's job). We decide where it landed — which competence lane it occupies, relative to the lane it was contracted for. That is a smaller question, and a decidable one.
The line is clean. WHETHER the work is good — correct, wise, true, what-you-really-meant — is undecidable in the deep sense: it runs into Goodhart, the is-ought gap, and the value-loading problem. WHERE the work landed — which actor-and-patient region of a fixed competence lattice it occupies — is decidable: it is a placement of two finite texts on a finite grid, recomputable by anyone offline, with no model in the loop. We insure and price the first kind of fact. We leave the second to the language model and the human, by design. The brain surgeon doing plumbing is not a quality error — the sutures may be immaculate. It is a located error: the patient-coordinate moved from a heart to a copper pipe, and the grid lights up exactly there.
Walk the canonical approaches and notice the one thing they share: each tries to certify the undecidable whole — is the system good, is it aligned with our values. Coherent Extrapolated Volition aligns to what humanity would want if we were wiser — beautiful, unmeasurable. Rules-based and Constitutional approaches specify the constraints — brittle, un-enumerable, Goodhart-prone. RLHF and preference learning train on human ratings — but the human is grading quality, the undecidable thing, and the model learns to hack the rating. Corrigibility keeps the system correctable — hard to verify. Mechanistic interpretability reads the weights — the most promising, but not yet decidable or scalable to autonomous speed. Five roads, one destination: a property no one can compute, so no one can guarantee it and no one can price it.
What we are doing is categorically different. We are not certifying values. We are measuring location — and the numbers actually tell us something, because the question is small enough to have an answer. This is not coherent extrapolated volition and it is not a rulebook. It is geometry — and the next section hands you the coordinate.
The reason none of the classic programs ship a price is the same reason they cannot ship a guarantee: they target a property that is undecidable in general. A decidable slice can ship both.
Here is the mechanism, and here is what it measured. We project an intent (the spec) and an execution (the work) onto a 144-node actor-and-patient lattice, walk the connectivity with a deterministic recursive definer walk, and read the Chebyshev (king-move) distance between where the work landed and where the spec asked for it. King-move distance matters: it weighs a diagonal shift exactly like an orthogonal one, so the agent must satisfy the constraint on every axis to count as in-lane — no "close on average" false positives.
The evidence is real and recomputable. On a sealed, blind, cross-domain held-out (financial covenants, legal conditionals, clinical dosing), off-domain work was rejected 10 out of 10, with a 0.90 separation between in-domain and out-of-domain. And we tested our own deepest worry — is this just string-matching dressed up? — with a controlled probe: the raw sensor rates a role-swap (same words, swapped roles) as 0.76 similar but a paraphrase (same meaning, different words) as 0.42. That tells us, honestly, that at the finest grain the sensor reads structure more than relation — and that limitation is the fence drawn exactly where it belongs: we certify domain membership (decidable), not within-domain relation (the model's job).
Two properties make the slice tradeable rather than merely interesting. First, it is hard to fool: the sensor is compression-based, so it reads the structure of the work, not its keywords. You cannot stuff the right words into off-domain work and slip past — off-domain prose with the vocabulary jammed in still does not compress like the domain. We measured that too. Second, it is safe to share: the proof is a signed receipt that carries only coordinates, a variance score, and one-way hashes of the ask and the work — never the proprietary work product itself. The platform proves the work happened in its lane without ever custodying a byte of it. A buyer, an underwriter, a counterparty can recompute the receipt offline and trust no one.
Decidable plus hard-to-fool plus shareable-without-leakage is the exact recipe a market needs in an underlying. That is why this slice can carry a contract.
Here is what it enables you, the reader, to do. Because lane-membership is decidable, signed, recomputable, and low-basis-risk — the trigger is the peril, not a proxy — you can write a contract on it. The first contract is insurance: it pays when the agent leaves its lane. The second is an option: you can buy the right to bet that a given agent, or a given competence authority, stays in-lane and completes in-domain tasks. This is not "skin in the game." Skin in the game aligns incentives by exposing the actor to downside. An option on competence is a tradeable instrument written on a decidable fact about an autonomous worker — a fundamentally new kind of asset. You are not trusting a reputation; you are pricing a recomputable boundary.
A competence authority — a model, a team, an agent fleet — becomes something whose in-lane completion is a priceable underlying. Buy options on it the way you would on any underlying with a measurable, bounded survival curve.
The pricing is not hand-waving; it falls out of the receipts. From a live ledger of attestations, the breach rate (the share of work that left its lane) lands near 13 percent with a tight confidence interval, the semantic volatility prices the load, and the premium comes out as base times the upper-bound breach frequency times one-plus-the-volatility-load — a number tight enough to write a policy on. Now scale it. Every autonomous agent task is a candidate contract. If even a fraction of the global market for delegated knowledge work routes through agents, and the in-domain, decidable fraction of that work is what becomes insurable and optionable, the underlying is the agent-labor economy itself — and a derivatives layer on agent labor is, in principle, larger than the labor.
The honest bound: the premium today is calibrated against our own in-lane ledger, not yet against blind held-out folds — that out-of-sample calibration is the next milestone, and it is the difference between "priced from our data" and "priced from work the model never saw." We name that inch rather than hide it. The point stands: a fair and bounded market on competence is not a metaphor. The number exists.
A market only deserves to be enormous if it is fair and bounded. The fence — domain membership, decidable, hard-to-fool — is what makes it bounded. The honesty about what we do not price is what makes it fair.
So what does it say for alignment, if this works — and it is a claim you can run experiments on, which is the whole point? It says the alignment problem has a labor that is now addressable: a slice of it — keeping an autonomous worker inside the competence it was assigned — is decidable, measurable, shareable, and tradeable, and that slice happens to contain the dangerous failure mode (competent action in the wrong domain) while explicitly excluding the undecidable remainder. And addressing only the decidable slice is not a half-measure — that is the move, not a shortfall. The undecidable remainder is not ignored; it stays where it belongs, with the human and the model. We carved off the part that is both dangerous and decidable — the agent in the wrong domain, the failure that bankrupts you — and made exactly that part computable and priceable. Solving a well-chosen piece is how every hard problem has ever gotten solved. Alignment stops being a single philosophical mountain and becomes, at least in part, an engineering field with a price. That is the growth: from coherent extrapolated volition you can only argue about, to a coordinate you can compute, contract on, and invest in.
For you, the reader — whether you build agents, underwrite them, or invest in the field — this is the moment a piece of "is the AI safe" turns into "here is the receipt, recompute it yourself, and here is the option." You become a participant in the first market where competence is the underlying. The physics is showing through clearly enough to bank on: not the whole of alignment, but a real, bounded, tradeable piece of it — and that piece is the one that bites.
Why no alternative, why not too-good-to-be-true, why nobody else. Why no alternative: the classic programs target the undecidable whole, so they cannot ship a guarantee or a price — a decidable slice can ship both (the boundary is Rice's theorem; two fixed artifacts on a finite lattice sit below the Turing line where Rice never reaches). Why not too-good-to-be-true: we measured the limitation ourselves and fenced it — we certify domain, not quality, and we say so. Why nobody else: the move requires giving up the prestige claim (we solve "good") for the humble, decidable one (we locate "where") — and pairing it with an oracle-not-host receipt so it can be shared without leaking IP. Foundational references: Yudkowsky on Coherent Extrapolated Volition (2004); Soares, Fallenstein, Yudkowsky, Armstrong on corrigibility (2015); Christiano et al. on deep RL from human preferences (2017); Bai et al. on Constitutional AI (2022); Rice's theorem (1953) on the undecidability of non-trivial semantic properties.
Run it yourself, and recompute every number offline, no model in the loop. npx thetacog-mcp attest-demo runs the placement and the insurability boundary; npx thetacog-mcp semantic-probe shows where the fence sits (the role-swap test); npx thetacog-mcp premium prices the peril. The deeper argument for why a reef-placement is recomputable and a model's verdict is not lives in Related: The Rice's Theorem Checkmate, and your competence coordinate has its own home at thetadriven.com/pixel.
The competence pixel is your coordinate. Being in it is being on-target, and that is the insured, optionable state. Your competence is no longer a black box — it is a receipt anyone can hold you to, and a market anyone can join.
→ Find your pixel: thetadriven.com/pixel — see the boundary, run the receipt, and claim the coordinate where where you are equals what you mean. That coordinate is the insured, optionable state, and the home of the competence market.
