Unscarcity
Sign in for free: Preamble (PDF, ebook & audiobook) + Forum access + Direct purchases Sign In By signing in, you get our daily podcast email. Unsubscribe any time.

Unscarcity Research

The Verifiability Premium: When Proof Replaces Credentials

Ten long-open problems solved for about $2,000, then Fermat's Last Theorem proved in Lean: where proof is cheap, credentials stop selling trust.

15 min read 3427 words Updated October 2026 /a/verifiability-premium-trust-collapse

Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

The Verifiability Premium: When Proof Replaces Credentials

The durable question for your profession is not whether a machine can do the work. It is whether anyone can check the answer without taking your word for it.

The $2,000 Receipt

On August 1, 2026, OpenAI published a paper titled Ten Advances in Mathematics and Theoretical Computer Science. Ten results, each of which, in OpenAI’s words, “resolves or makes substantial progress on a long-standing open problem,” produced by an internal version of a then-unreleased model called Astra, the one that shipped as GPT-6 Astra on September 3 to a first set of customers. Among them: a disproof of Connes’ rigidity conjecture in operator algebras, and the first improvement since 1978 to the general exponent for how densely you can pack spheres in high dimensions.

The capability is the smaller half of the story. Labs announce breakthroughs constantly, and the standard response is to wait for someone credentialed to look at it and tell you whether it holds. That is not what happened here. Each argument was then formalized into a certificate for Lean, a formal proof checker, and the certificates sit in a public GitHub repository. The result does not ask for your trust. It hands you a file and invites you to run the checker yourself.

OpenAI put the price on it: the tokens needed to find all ten solutions “would cost roughly $2,000 at Sol API rates.” Noam Brown, one of the researchers behind the underlying reasoning work, volunteered the caveats on X, which is more than most announcements manage. They tried other major problems and failed. “Sadly, no Millennium Prize Problems (yet).” And “we didn’t spend a lot on each problem,” so $2,000 is the bill for the wins, not a price list for mathematics.

Hold the number anyway, because a floor is still a floor. Ten results on problems the field had left open for years arrived for roughly the cost of a used car, and arrived pre-verified.

What a Credential Was Always For

A credential is not a measure of knowledge. It is a trust technology, and like every trust technology it exists because verifying the underlying thing directly was too expensive.

You do not read your accountant’s work. You cannot evaluate your structural engineer’s load calculations. You have no way to independently confirm that the radiologist saw what they say they saw. So instead of checking the output, you check a proxy: a degree, a board certification, a license, a track record, a firm’s letterhead. Society built an entire apparatus of examinations and professional bodies for one reason, which is that direct verification of expert work was historically harder than vetting the expert.

That is a conditional arrangement, not a law of nature. It holds only while verification stays expensive.

The Substitution Threshold defines the moment when the cheapest reliable provider of a service stops being human, and the word doing the work in that sentence is “reliable.” A credential is how reliability gets established when you cannot inspect the work. Take away that constraint, and the credential loses its job.

A Lean certificate is what taking away that constraint looks like. It reduces “do I trust the entity that produced this proof?” to “does the file compile?” The second question costs a few seconds of CPU time and does not care whether the author was a Fields medalist, a graduate student, or a model with no release date.

Two Costs, One Ratio

Every piece of knowledge work has two prices attached: the cost to generate an answer and the cost to verify it. Almost everything about how a profession is organized falls out of the ratio between them.

When generation is expensive and verification is also expensive, you get guilds. Credentials, licensure, reputation, long apprenticeships, and the whole social machinery of vouching. This is medicine, law, structural engineering, auditing.

When generation is cheap and verification is cheap, you get commodity markets. Nobody asks about a courier’s credentials, because you can see whether the package arrived.

AI has been collapsing the generation cost across the board for four years. That alone does not dissolve a profession, which is why the doctor and the accountant are still employed while a model can already out-read both of them. What the Astra paper demonstrates is the second collapse, in the domains where it applies: verification cost falling to near-zero and falling into machine hands.

Where both costs collapse together, price is the only variable left. There is nothing else for a credential to do.

Which Work Has a Certificate

This gives you a sorting rule far more useful than “is my job creative?” Ask instead: does my output have a checker?

Work that already ships with something like a certificate:

  • Code, checked by a test suite, a type system, a fuzzer, a build that either passes or does not. This is why software felt the shock first, and why the developer job ladder lost its bottom rungs before anyone’s ladder did.
  • Formal mathematics and cryptography, checked by proof assistants like Lean.
  • Tax positions, checked against statute, precedent, and eventually an audit.
  • Structural load calculations, checked by simulation and by physics that renders its verdict without a hearing.
  • Actuarial and financial models, checked against realized outcomes and regulatory stress tests.
  • Drug candidates, checked by assays and trials, slowly and expensively, but mechanically.

Work with no checker at all:

  • Deciding which problem is worth solving. Nothing verified that a human chose all ten of Astra’s problems.
  • Judging whether a client should settle or fight, when both branches are defensible.
  • Telling a patient what a diagnosis means for the life they actually wanted to live.
  • Setting the objective function in the first place, which is where Goodhart’s Law eats organizations alive.
  • Being the person who can be held responsible when the answer turns out to be wrong.

That last item is a separate firewall, and it is worth being precise about how it differs. The liability gap argues that a professional license sells accountability, a named human the system can punish. Verifiability and accountability are not the same shield. A Lean certificate proves that a theorem follows from its axioms. It says nothing about whether the theorem was worth proving, and there is no one to sue if it was not.

So a profession has two things holding the line, and they fail independently. Verifiability failing means your work can be checked without you. Accountability failing means nobody has to answer for it. Where the first collapses and the second was never required, the profession clears out fast. Where the first collapses but the law still demands a suable human, you keep the human and shrink the job to a signature.

The Field That Went First, and Said So Out Loud

Mathematics is the interesting case precisely because it built its own verifier and then had to live with the consequences.

The profession saw this coming. The Leiden Declaration on Artificial Intelligence and Mathematics was published on June 2, 2026 out of a workshop at Leiden University and endorsed by the International Mathematical Union. It warns that fast-moving developments put the field’s system of review under pressure, “jeopardizing our ability to implement traditional standards for the correctness, transparency, and independent verifiability of proof.” Current automated techniques, it adds, “can produce plausible but unreliable (or even incorrect) arguments which are difficult to distinguish from correct mathematical proofs.” It does not ask for a ban. It asks the community to make its values explicit before the norms get set by whoever ships fastest. It had passed 3,000 signatures by late July and 4,300 by late October. Terence Tao endorsed it wholeheartedly, singling out its call for mathematicians to participate in public discourse, and said the age of AI had made it important to spell out the community’s goals and values, which had long been left implicit.

Timothy Gowers was part of the workshop and declined to sign, not out of disagreement so much as discomfort with how confident several of its assertions were. His own reckoning, in the same post, is more personal. Twice, GPT-5.6 Pro one-shot a problem he had cared about and thought hard about, prompted by younger collaborators with his approval; it felt “very strange and not particularly pleasant to have the rug pulled out from under my feet like that,” while he was still “quite pleased to see the problems solved.” His larger fear is not about credit. It is that after a decade or two the literature has been “vastly expanded, but there is no corresponding community of human experts who have a shared understanding of parts of it.” The unease sharpened in September: on September 11, twenty-five Fields Medalists, Tao among them, published a joint declaration on mathematics and AI, which the Economist headlined “Top mathematicians are outraged by OpenAI’s methods,” and Gowers stayed off that list too, explaining why six days later.

Not everyone experienced it as loss. Abhishek Saha, a mathematician at Queen Mary University of London, wrote in July that in his area frontier models are already “at least as good as a solid and indefatigable PhD student.” By October he was handing the model a bare sketch of a strategy and trusting it to fill in the details: “Increasingly I find myself taking on the role of conductor, rather than doubling up as orchestra.” That is the same restructuring the book describes in The Solo Unicorn: one human holding taste while the labor underneath goes elastic.

Two responses, one field, same month. Both are correct. Conducting is a real job, and there are fewer chairs in a conductor’s career than in an orchestra’s.

September: What the Certificate Checks, and What It Cannot

Five weeks after the Astra paper, the two biggest proof announcements of the year landed four days apart, and between them they draw the boundary of the verifiability premium more sharply than any argument could.

On September 4, Anthropic reported that dozens of Claude agents, running largely unattended for eleven days, had produced the first complete machine-checked proof of Fermat’s Last Theorem. The tally: 13 million lines of Lean, 29,500 intermediate theorems in the final proof, more than five times the size of Mathlib, about six billion output tokens from an internal research model roughly comparable to Claude Fable 5.1. Human input was limited to occasional high-level instructions. Kevin Buzzard of Imperial College, who is funded by the EPSRC to formalize the same theorem, compiled the code base himself, ran the comparator, read every line that was not a definition or a proof to make sure nothing was hidden in the machinery, and reported that it checks out. His post is titled “Anthropic has beaten me to it.” He also filed the caveats a verifier cannot. Mathematically, as he put it in the comments two days later, “the work tells us nothing new mathematically, so mathematics is unchanged by this work.” The proof, by Anthropic’s own footnote, “is likely much longer than it needs to be.” And it covers exponents from 17 upward, leaning on earlier formalizations for the rest.

Notice what made an eleven-day unattended run trustworthy. Not the agents’ confidence, and not the lab’s word. The Lean kernel checked every step, and a human with the standing to reject the result ran the final check anyway. That is the two-part arrangement this article has been describing: the checker replaces the credential for the question “is this proof valid,” and the accountable human stays for the questions the checker cannot hear.

On September 8, OpenAI said that roughly 10,000 of its agents had, in about 88 hours, produced “an analytical proof and a Lean formalization that an initially smooth fluid at rest can develop a singularity in a finite time” under the Navier-Stokes equations, one of the Clay Millennium Prize problems; translating the argument into Lean took a further 17 hours. The certificate compiles. That is exactly where it stops being useful. The proof needs an external force pushing on the fluid, which Charles Fefferman’s 2000 problem statement allows in its two breakdown cases and sets to zero in its two existence cases, while the unforced problem, the one the field has spent decades on, is untouched. Three mathematicians then showed that the blow-up disappears once the force is removed. “The Clay problem is settled, but the main problem for the Navier-Stokes equations is not,” is how Luis Silvestre of the University of Chicago put it.

Priority is disputed too. Tristan Buckmaster and Levent Alpöge had posted their own forced blow-up proofs, built on Diego Córdoba and Luis Martínez-Zoroa’s technique, hours before OpenAI’s announcement. OpenAI says its run began on September 1 after “hearing a rumor” of their progress. Buckmaster asked whether OpenAI’s models had access to the pair’s Codex sessions, which OpenAI emphatically denies. The experts who read the paper for Scientific American in October found the similarities superficial and the approach different. The Clay Institute says only that the problem has “apparently been settled”; its president, Martin Bridson, calls the evaluation “deliberately unhurried”, the rules require a peer-reviewed publication and two further years of acceptance before a committee even convenes, OpenAI says it will not claim the money, and the Institute’s problem page still carries Navier-Stokes as open.

Then the certificate itself went under the microscope. On October 6, three Cambridge mathematicians, Alexander Bastounis, Fabian Circelli and Anders Hansen, posted a paper arguing that checking an AI’s own Lean translation “may offer no confidence in the original” natural-language argument, and showing “that the formalised Lean proof does not correspond to the NL proof of blow-up of solutions to the Navier-Stokes equations.” NPR put it plainly on October 21: the AI “made unexpected changes to its work when it translated the proof into Lean,” and the statement in the written proof, Hansen says, is stronger than the one Lean checked. The mismatch does not make the result wrong. It means the thing that compiles is not quite the thing that was claimed, the one failure a checker cannot report, because nobody asked it the right question. The same NPR piece follows the humans grading the repository of more than 370 further AI-written results OpenAI released in October. Dana Moshkovitz found the paper on her own Unique Games Conjecture “looks right” but “not made for human consumption.” Daniel Litt found a subtle error, “like a sign error, basically,” in an algebraic-geometry paper, which, as NPR put it, “means the whole paper is wrong.” Anthropic and OpenAI have since committed about $5 million each to help researchers evaluate AI-assisted results, which is what it costs to keep a human at the end of the chain.

Put the two side by side. Both proofs carry a Lean certificate. One is accepted; the other is disputed on every axis that matters. The difference is not in the checker. A Lean certificate answers exactly one question: does this proof establish this statement from these axioms. It cannot tell you whether the statement is the one that was asked, whether the hypotheses were chosen to make the answer come out, or whose idea the approach was. Those are the items on the no-checker list above, choosing the problem and answering for the result, and the September dispute is what it looks like when a machine-verified answer arrives with both of them unresolved, and October added a third: whether the statement the checker verified is even the statement in the paper. Terence Tao, writing about the human proofs that landed the same week, gave the line to keep: solving these problems “is only a proxy goal for the primary goal of developing mathematical understanding and insight.” The certificate covers the proxy, when it covers it. The premium moves to everything it does not.

The Unscarcity Reading

The book’s argument has never been that AI will eventually get good enough to threaten employment. It is that the Labor Cliff arrives through accounting, one task at a time, wherever the arithmetic clears. The verifiability premium tells you the order of the queue.

Three things follow.

First, the cliff reaches credentialed work earlier than the standard story predicts. The reassuring version of automation says routine work goes first and expertise is safe. The verifier says otherwise. Research mathematics has an unusually good checker, so research mathematics got a public bill attached to a list of long-open problems before most clerical work did, and then watched its most famous theorem formalized in eleven days. Meanwhile the employment data shows displacement concentrated among the youngest workers, the ones whose output was most easily checked by a supervisor. Checkability, not seniority, is the exposure.

Second, the surviving premium is on taste and on answerability. If the machine can search hard with a verifier attached, the scarce human contributions are choosing which search to run and standing behind the result. That is a smaller number of jobs than the profession currently supports, and it is a genuinely different skill from the one most professional training optimizes for. It also demands the thing human-in-the-loop design keeps insisting on: a minimum layer of accountable people who did not delegate the part that cannot be delegated.

Third, this is exactly why the floor has to come before the transition, not after. A world where verified answers cost $200 apiece is a spectacularly good world to live in, provided your ability to eat does not depend on being the person who used to sell them. The Foundation exists to decouple survival from that dependency. The Frontier exists so the people freed from selling verified answers have somewhere to put their ambition. The alternative is a decade in which the cost of expertise collapses while the cost of rent does not.

The Question to Ask About Your Own Work

Forget whether AI can “do your job,” which is a question nobody can answer and everybody enjoys arguing about. Ask the narrower one:

If a machine produced my output tomorrow, how would anyone know whether it was right?

If the answer is “run the tests,” “check it against the statute,” “simulate the load,” or “compile the proof,” then the trust technology you were selling has a replacement in the field already, and price is the only remaining argument. Plan accordingly, and move toward the parts of the work that no checker covers.

If the answer is “you would have to trust the person who made it,” you have more time. Not because the machine is worse, but because nobody has built the verifier yet.

And if the answer is “nobody would know for years, if ever,” check whether that is expertise or whether it is something the profession has been able to avoid measuring. Those look identical from the inside, right up until someone attaches a checker.

The machine did not take mathematics. It attached a price tag to a sentence mathematicians had never had to price before: “I could have done that eventually.” Every profession with a verifier is next in line for the same invoice. Read the book or start with the preamble.

Share this article:

Share Share Share