Published on: June 30, 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 bounded-function story is the most expensive misunderstanding in AI assurance — expensive precisely because it arrives disguised as good news. The argument runs like this: a large language model is not really one of those undecidable Turing machines, it is just a bounded function, finite in and finite out, so you can finally inspect it, score it, and certify that its work is good. If that were true the whole "AI is uninsurable" problem would dissolve, and every evaluation suite on the market would be selling something real. It is half true — and the half that is false is the half your liability lives in. So we are going to pin the thing to the table and find the exact seam where the argument quietly switches specimens on you. The true half describes a machine nobody deploys. The false half is the machine you actually run.
If you believe the bounded-function story, you will buy the one product that cannot exist: a general decider for whether an AI did a good job. Seeing through it is not academic. It is the difference between budgeting for an answer and budgeting for an oracle.
Give the skeptic their strongest ground, because the honest part of their case is genuinely honest. A transformer with frozen weights doing one forward pass is a fixed-depth circuit. It has no loop, no growing memory, no way to extend its own work — it is the same shallow computation every time, the kind theorists file under a constant-depth class they call TC-zero. That is a bounded function in the truest sense, and yes, in principle you could enumerate what it does. So far the skeptic is right, and you should say so out loud. The sleight is not in this sentence. The sleight is in the next one, where they hope you will not notice that no one on Earth deploys a single forward pass.
The model you put in production does not answer in one breath. It writes a token, appends it to its own context, reads the longer context back, and writes again — over and over. That feedback is autoregression, and it is not a detail. A Turing machine is nothing more than a small fixed controller plus a tape it can keep writing to — and a model that appends each token to its own context and reads it back is that controller, writing to that tape. Hand it a scratchpad and tell it to think step by step, and that loop reaches full Turing completeness — this is a proved result, not a metaphor. The bounded-function argument was describing a freeze-frame of a machine whose entire nature is the motion between frames. You cannot bound a river by photographing it.
Chain-of-thought is not a prompting trick. It is the model using its own output as working memory. The instant a system can read what it just wrote, the tape is unbounded in every way that matters, and the bounded-function freeze-frame is gone.
Suppose you do not even grant the loop. The agent that holds your balance sheet still calls tools — a code interpreter, a shell, a database, a browser, another program. Every one of those is a Turing machine in its own right. A system that can invoke a Turing machine and decide what to do next based on its output is a Turing machine — it has borrowed the unbounded tape it needed. The moment it runs one line of code you did not write, the limit of what it can do becomes the limit of what a computer can do — which is to say, none. This is the part of the argument with no escape hatch: undecidability is contagious upward. A body that picks up a Turing-complete instrument inherits the undecidability of everything it can do while holding it. There is no version of a useful agent that never reaches for the scalpel — the reaching is the whole reason you hired it.
Here is the last refuge, and it deserves a real answer because it is technically correct. Any physical computer has finite memory, so strictly it is a finite-state machine, and finite-state machines are decidable. True. Now count the states. That "finite" machine has more configurations than there are atoms in the observable universe, by a margin so large the comparison stops meaning anything. The decision procedure exists and will finish its enumeration a few hundred orders of magnitude after the last star burns out. Decidable-in-principle and uncomputable-in-this-universe are the same fact wearing different clothes, and you cannot write a premium on a procedure you can never run. Tell an actuary "the answer is knowable, you just need a computer larger than the cosmos" and they reach for the exact same pen as the one told it is undecidable: they stop trying to decide the outcome and start measuring how often it breaks.
"Decidable in principle" is the most expensive phrase in AI assurance. It is technically true and operationally worthless, and an entire category of products is priced as if the gap between the two does not exist. It is the gap.
The most sophisticated version of the dodge says the training itself constrains the model down below a Turing machine — that all those weights fence it into something tame and decidable. The honest shape of a language model is a Turing machine plus a paradigm: a learned prior over which computations are worth running. A prior bends the odds toward some behaviors and away from others. It never lowers the ceiling. To pin the system below the Turing line you would have to prove it can never express a Turing-complete computation — and the model expresses them every day, the moment it writes a loop in a scratchpad or runs one in a tool. Training tells the machine what it usually does. It cannot tell the machine what it cannot do. So the constraint claim is not merely weak. It is false in the lab, on the table, in front of you.
Here is the line that turns all of this from a debate into a decision you can act on, and it is the opposite of what the evaluation vendors are selling. The law: if any reachable part of a system is Turing-complete, its non-trivial semantic properties are undecidable. That is Rice's theorem, and it is why "did the AI do a good job" has no general decider and never will, no matter the budget. The fool's move is to claim you beat the theorem. We do not. We sidestep it: we never ask whether the work was good, because that is undecidable; we ask whether the agent stayed inside the lane you authorized — and where it landed is a decidable, sub-Turing fact you can sign and recompute on your own machine.
Concretely, that changes what you can do with the exposure. Today you delegate work to agents you cannot fully verify and carry the result as an unpriced unknown. A signed record of where the agent stayed turns that unknown into a line you can act on: a recomputable fact you can show a board, attach to a contract, or hand a carrier to actually transfer the risk. You are not promised the work was good — you are given the one thing that was missing, a number for where it landed and a measured history of how often that boundary breaks. That is the whole distance between a risk you can only worry about and a risk you can price, exclude, or insure. And because the boundary recomputes to the same number on your own hardware, none of it asks you to trust us.
The machine was never going to hand you decidable quality. It was always going to hand you a decidable lane. Stop buying the quality. Buy the lane.
You do not have to take any of this on faith — that is the point of choosing the decidable slice. The boundary recomputes the same number on your hardware, with no model in the loop. Run it yourself: npx thetacog-mcp attest-demo. A claim you can re-derive without the vendor is the only kind an underwriter was ever able to price.
None of this is pessimism. The meaning of an AI's work stays undecidable — not because we failed to measure it, but because the theorem says no one can — and that is exactly what makes the decidable part worth owning: not whether the work was good, but where it stayed, signed and recomputable on your own machine. Stop paying for the answer no theorem allows. Start pricing the lane that is already there. If you carry AI risk for a living, the people drawing that boundary are gathering at one table. See where the lane gets drawn — thetadriven.com/dinner.
