Abstract
Two debates in the philosophy of mathematics have stalled: whether computer-assisted proofs that no one can survey are proofs in the traditional sense, and whether probabilistic proofs are epistemically inferior to deductive ones. I argue that both rest on an unexamined assumption, proof-level predication: that surveyability, transferability, reliability and explanatory power are properties of a proof taken as a whole. Against it, I defend the Locus Thesis. These properties belong primarily to the components of a warrant architecture (a specification linking the intended claim to an object-level statement, a justificatory object, a checking procedure, a trusted base, an error profile, an allocation of epistemic functions, a set of independent replication routes, and a selection profile), and they hold of the whole only through explicit composition rules. Read through modal epistemology, the thesis is a principled response to the generality problem for mathematical reliabilism in computational mathematics: under conditions that can be stated precisely, the method that confers warrant is the acceptance procedure, and the process that generated the proof is screened off. Several connected consequences follow. A Randomness Location Principle, derived from sensitivity and safety, splits the probabilistic-proof debate: randomness bears on warrant only where it sits in acceptance. Surveyability is reconstructed as layered auditability, and the novelty of AI mathematics is located precisely: when generation is cheap, doubt migrates to specification and selection. OpenAI’s release of 722 machine-produced manuscripts on 6 October 2026, some formally verified and some not, is analyzed through a notion of release-level warrant, with an interventionist audit procedure, and the public controversy it provoked is shown to divide along the components of the architecture. Explanation is shown to be a separate function that warrant does not supply, and eight objections are answered.
1. Introduction
On 6 October 2026, OpenAI made public a repository of 722 mathematical manuscripts, organized into 372 result families, produced by an unreleased internal model that had been posed approximately 4000 problems; according to the repository, each result used “three hours of ChatGPT Pro thinking compute on average” [1]. The repository’s own description is careful: the collection “includes results at different stages of verification”, “many, but not all, of the manuscripts have been formalized” in the Lean proof assistant, and “some of the unformalized results could have issues” [1]. Fifteen days earlier, an independent Advisory Group on Mathematics and Artificial Intelligence hosted at the Institute for Advanced Study had been announced [2]; a week before the release, it had published guidelines asking AI laboratories to stop testing advanced problems on proprietary models and, where they nonetheless released results, to disclose how problems were selected and how many comparable problems their models had failed to solve [3]. Within a day, the release had divided the mathematical community: it was welcomed as “great for mathematics”, denounced as “a demonstration of power”, and read as the end of a way of doing mathematics centered on raw problem solving [4,5,6].
Fifty years separate this release from the summer of 1976, when Appel and Haken announced a proof of the Four-Color Theorem whose reducibility component no human being could check by hand [7,8]. The philosophical reaction to 1976 framed a question that still organizes the field: is a proof that cannot be surveyed a proof in the traditional sense [9]? A parallel debate, opened by Fallis, asked whether arguments that establish a conclusion only with high probability are excluded from mathematics for any good epistemic reason [10,11,12]. Both debates are mature, and neither has converged.
This article argues that the two stalemates share a root, and that the root matters more for the October 2026 release than it did for 1976. Both debates assume, usually without saying so, that the properties under dispute (surveyability, transferability, reliability, a priori status, and explanatory power) are properties of a proof considered as one object. I call this assumption proof-level predication (PLP). It is a reasonable idealization for a short paper proof read by a competent colleague. It fails for every case that generated the debates, and it fails decisively for machine-produced mathematics at scale. The central claim of the paper is the following.
Thesis 1
(Locus Thesis). The epistemic properties traditionally predicated of proofs are primarily properties of the components of a warrant architecture: a specification, a justificatory object, a checking procedure, a trusted base, an error profile, an allocation of epistemic functions, a set of replication routes, and a selection profile. A property holds of the architecture as a whole only when an explicit composition rule licenses the inference from its components.
The thesis is contestable. A defender of PLP can hold that “proof” names a kind whose credentials are irreducibly global, so that decomposition changes the subject; Section 11 answers eight versions of this resistance. The thesis also has a deeper motivation than descriptive convenience. Reliabilist and modal accounts of knowledge must say which belief-forming method is being evaluated, and they face the familiar difficulty that a process can be described at many levels of generality [13]. In mathematics, PLP individuates the method as “the proof”, which is too coarse, since it lumps together the process that found an argument and the procedure that accepted it. The Locus Thesis supplies a principled individuation: the method that confers warrant on a mathematical claim is its acceptance procedure, and the generation of the proof belongs to that method only when acceptance does not screen it off. Section 4 develops this point, and most of the paper’s results follow from it.
The problem the paper addresses is therefore a general one in the epistemology of computational environments: when the production, checking and understanding of a mathematical result are carried out by different human and computational agents, what exactly is warranted, by what, and for whom? I treat it as a problem in computational epistemology, in the sense in which the epistemology of science had to be extended once computation became constitutive of how knowledge is produced and not only of how it is applied [14,15].
The Locus Thesis and the acceptance-individuation argument developed in Section 4 are the central philosophical proposals. The four connected consequences below examine their implications in different settings; they are not four independent theories.
- 1.
- The two stalemates are partly verbal: in each, the disputants are right about different components of one architecture (Section 6). The probabilistic-proof debate yields a Randomness Location Principle, derived from sensitivity and safety: randomness bears on warrant only where it is located in acceptance.
- 2.
- Surveyability by a single reader should be replaced by layered auditability, which states what a community must be able to do with each component and how component audits compose (Section 9).
- 3.
- What is new in AI mathematics can be located precisely. Machine participation in proof is old. What changes is an economy: when generation becomes cheap relative to checking, doubt migrates from the justificatory object to the specification (does the formal statement say what was meant?) and to the selection profile (how many candidates were produced, and by what rule were some retained?). The October 2026 release calls for a notion of release-level warrant and for an interventionist audit procedure to estimate it, and the controversy that followed it divides, position by position, along the components of the architecture (Section 7).
- 4.
- Warrant and explanation come apart in a principled way. The structural tests by which the author’s earlier work on xenoepistemics evaluates non-human knowledge producers [16] are tests of generators; the demand for proofs that humans can navigate, compress and reuse is a demand on a different function. Keeping the two apart clarifies what AI “co-mathematics” should and should not be asked to deliver (Section 8).
Some associated questions, especially mathematical authorship, editorial governance and AI co-mathematics, are treated as bounded applications. The discussion indicates why the framework bears on them without claiming to supply a complete theory of responsibility, institutional governance or mathematical collaboration. These topics merit separate investigation.
Table 1 states how these claims relate to existing scholarship on the four themes that discussions of AI in mathematics usually run together: agency, authorship, opacity and understanding.
Table 1.
Relation of the paper’s claims to prior scholarship.
- Method and scope.
The method is historical-philosophical. Historical episodes are used as evidence in a specific sense: each was chosen because in it a property that PLP treats as global can be seen migrating from one component of an architecture to another while the proposition stays fixed (Section 5). The same criterion explains why the evidence begins with deductive uses of probability: the probabilistic method, randomized algorithms and computer-assisted proofs belong to one narrative because each relocates randomness or computation to a different component of the same structure, and the contrast between those locations is what the Randomness Location Principle exploits. The article is not a survey of automated reasoning. It omits systems that improved performance without changing the structure of warrant, and it introduces technical notions only where an argument depends on them.
2. Proof-Level Predication and Two Stalemates
2.1. Surveyability
Wittgenstein’s remark that a proof must be surveyable (übersichtlich) made surveyability a candidate criterion of proofhood [28]. Tymoczko turned the remark into a challenge after the Four-Color Theorem. The reducibility of the unavoidable configurations had been established by computer, so no mathematician had surveyed the whole argument; the warrant for the theorem included the empirical proposition that particular machines running particular programs had produced particular outputs. Tymoczko concluded that accepting the theorem changed the concept of proof and introduced experiment into mathematics [9].
The replies took several forms. Teller argued that surveyability had never been a requirement on proofs as such, only on our coming to know them, and that reliance on others’ checking was already ordinary [29]. Detlefsen and Luker argued that ordinary proofs already contain empirical premises (that a long hand calculation was performed correctly, for instance), so the computer introduced no new kind of evidence [30]. Burge argued on general grounds that reliance on computers in a proof need not make the resulting knowledge a posteriori, because the entitlement to rely on a properly functioning source of information can be a priori, as reliance on an interlocutor can be [31]. Bassler later showed that the debate had worked with an impoverished notion of surveyability and distinguished surveying the local steps of a proof from surveying its global structure [32].
The disputants share an assumption. Tymoczko says that the proof is not surveyable and that its warrant is partly empirical. His critics say that the proof is deductive and that its warrant is no more empirical than that of long hand-checked proofs. Each side predicates one epistemic status of one object. Bassler’s distinction is the first step away from this, but it remains a distinction between two ways of surveying a single proof.
2.2. Transferability and Probabilistic Proof
Fallis asked why mathematicians accept proofs that are long, gappy or partly carried out by computers while rejecting probabilistic proofs, such as the Miller–Rabin verdict that a number is prime, whose probability of error can be made smaller than that of any human-checked proof [10,33]. He argued that no epistemic reason supports the asymmetry. Easwaran replied that acceptable proofs have a property that probabilistic proofs lack, transferability: a suitably expert reader can come to know the conclusion from the proof alone, without relying on testimony about how the proof was produced. A probabilistic proof fails this test because the reader must trust that the random choices really were random and were not selected to make the test pass [11]. Jackson answered that randomized arguments are transferable after all, since any reader can rerun the procedure with randomness of her own [34]. Hamami offered a different reason for rejection: knowing a conclusion on the basis of a probabilistic proof would be relevantly like knowing a lottery proposition on statistical grounds, which most epistemologists deny is possible [12]. Paseau, from the other side, argued that non-deductive evidence can suffice for mathematical knowledge [35].
The shared assumption is again PLP. “Is the probabilistic proof transferable?” presupposes one object, the probabilistic proof, of which transferability is true or false. Section 6 argues that the disputants are describing different components of one procedure.
2.3. The Assumption Stated
Proof-level predication (PLP). For each epistemic property relevant to mathematical acceptance (surveyability, transferability, reliability, a priori status, explanatory power), there is a determinate fact whether a given proof of P has , and the acceptability of P on the basis of depends on whether has .
PLP fits its paradigm: a short informal argument, read by a competent colleague who checks every step, in which the production of the argument, the object checked and the act of checking are all realized in a few minds. Every case that generated the debates breaks this unity, and the next section introduces vocabulary for what replaces it.
3. Warrant Architectures
3.1. Definition
Let P be an intended mathematical claim. A warrant architecture for P is a tuple
with the following components.
- S, the specification: the representational link between the intended claim P and an object-level statement (a Lean theorem statement, a propositional formula, a list of configurations, a finite predicate) on which the remaining components operate.
- O, the justificatory object: an informal derivation, a finite case list with code, a primality certificate, a machine-checkable refutation, a proof term, or an explicit construction.
- C, the checking procedure applied to O, with output .
- T, the trusted base: whatever must be correct for to hold (axioms, the proof assistant’s kernel, compiler, hardware, libraries, or sources of randomness).
- , the error profile of C: zero, one-sided or two-sided, with its bound.
- , the allocation of six epistemic functions to human and computational agents .
- R, the set of replication routes: materially independent reimplementations, re-proofs or reformalizations, as opposed to repeated runs of the same artifact.
- , the selection profile: the population of candidates from which O was retained, the process that generated them, and the retention rule.
Two components carry most of the argument about AI. S is often absorbed into the trusted base. Separating it records that statement fidelity is something no kernel checks, and that it is audited by different people (domain experts) from those who audit C and T. is degenerate in the classical paradigm, where a mathematician typically publishes the one proof she found; it becomes dominant when a system produces thousands of candidates.
The six functions in are: problem formation (selecting and stating a worthwhile question); discovery (generating conjectures, examples, representations and strategies); justification (constructing a derivation, certificate or error-bounded argument); verification (checking that the justificatory object satisfies accepted rules); explanation (compressing the result into concepts that support understanding and reuse); and stabilization (reviewing, publishing, teaching and integrating the result into a field). The point of is not to attribute agency in a metaphysically loaded sense. It records where the operations necessary for a result were performed, in the spirit of distributed-cognition analyses [17], and it makes visible what Friedman calls the role inversion of contemporary practice, in which mathematicians increasingly check, interpret and steer what machines produce [19].
3.2. Composition Rules
The Locus Thesis says that global properties hold only through composition rules. Three are needed below.
Error composition. Let be the probability that the specification is unfaithful (that does not express P), the probability that C accepts an O that does not establish , and the probability that some element of the trusted base fails in a way that matters. For any probability measure over these failure events, the union bound gives
The bound is elementary, and that is its point: a global error estimate is available only when each component has received one, and different agents typically supply them. A kernel that drives to zero does nothing for .
Weakest-link fidelity. If S is a chain of representations (a combinatorial problem, its propositional encoding, the encoding after symmetry breaking), fidelity of the whole requires fidelity of each link.
Independence for replication. An element of R adds warrant only to the extent that its failure modes are independent of those of the original route. Rerunning the same binary on the same hardware adds almost nothing; an independently written checker or an independent proof adds a great deal [36].
3.3. What the Framework Is Not
W is a typed description, not a numerical ranking, and it does not reduce proof to sociology. It is compatible with the mathematician’s criterion that a derivation accepted by a correct kernel, relative to a stated formal system, is a formal proof. Explanatory quality is not among its components: no function can be assumed to deliver understanding, since formally equivalent objects differ greatly in conceptual compression, and formalization sometimes creates understanding instead of merely consuming it [27]. Explanation is attached to an architecture through ; Section 8 returns to it.
For example, a stochastic search might discover a proof term that a sound kernel subsequently accepts. The search history affects how frequently such a term is found, whereas the kernel checks derivability of the stated proposition. Neither operation, by itself, establishes that the formal proposition faithfully expresses the original mathematical question, or that an entire selected collection has been adequately vetted. The distinctions therefore preserve the usual criterion of formal correctness while locating separate epistemic responsibilities.
4. Which Method Confers Warrant? A Modal Grounding
The Locus Thesis would be a classificatory convenience if it had no basis in the theory of knowledge. This section gives it one.
Modal conditions on knowledge evaluate a belief relative to the method by which it was formed. On Nozick’s sensitivity condition, a subject knows p only if, were p false, she would not believe p via the method she actually used [37]. On Sosa’s safety condition, a belief is safe if, in nearby worlds where she believes p via that method, p is true [38]; related safety-based treatments explain why lottery propositions are not known on statistical grounds alone [39,40]. Both conditions inherit the generality problem: the same token belief-forming process falls under types of very different reliability, and an account that does not say which type is relevant says nothing determinate [13].
In mathematics, the problem has a definite shape. Consider a mathematician who believes because a proof assistant accepted a proof term found by a stochastic search. One candidate method is “believe whatever the search outputs”. Another is “believe when the kernel accepts a term for ”. The two differ in modal profile. In worlds where is false, the search may output garbage, but a sound kernel accepts no term for a false statement. PLP does not choose between these descriptions, because it treats “the proof” as the unit. The Locus Thesis chooses, and the following principle states the choice.
Principle 1
(Acceptance individuation). When a justificatory object is retained only if an acceptance procedure C accepts it, the method relevant to the modal evaluation of a belief in is the acceptance procedure together with the specification and trusted base on which it depends, . The generating process belongs to the method only to the extent that some retained object is not screened by C.
The principle applies to mathematics the general idea that a method is individuated by what the subject’s belief is sensitive to: a belief formed by checking is sensitive to the outcome of the check, whatever produced the object checked. The final part of this section defends it against the natural alternative of individuating the method as the whole generating pipeline. Two consequences, both used below, follow at once.
First, if C is deterministic and sound relative to T, S is faithful, and acceptance screens off generation in the sense made precise below, a belief in P formed via is sensitive and safe whatever the generator does: in no world where P is false does C accept an object for . Randomness in generation changes whether a belief is formed (how often the search succeeds), not whether a formed belief is true. In the vocabulary of Section 7, generation affects the power of the procedure, and acceptance alone fixes its false-acceptance rate.
Second, if C has two-sided error ( in the sense that acceptance is compatible with falsity), acceptance alone cannot rule out a false conclusion. In cases where false-acceptance outcomes occur in the relevant nearby worlds, a belief formed through C may fail safety, much as a belief in a lottery proposition formed solely on statistical grounds may fail safety. This is the structural concern behind Hamami’s argument [12]: a very small chance of error does not by itself turn a probabilistic acceptance into the kind of warrant that a sound deductive check provides. The qualification concerning nearby worlds matters. A nonzero numerical error bound alone does not determine modal proximity; the safety verdict also depends on the method and the relevant counterfactual circumstances. The architectural point is that Hamami’s objection targets randomness in the acceptance procedure and its error profile, not randomness confined to the search that produced a subsequently checked certificate.
Why Not the Whole Pipeline? Screening-Off Conditions
The strongest objection to acceptance individuation is that it is too convenient. Reliabilists usually individuate a belief-forming process by its actual causal history, and the actual causal history of a machine-produced theorem includes the generator, whatever selected among its outputs, and the checker. Why should the epistemically relevant method not be this whole pipeline? If it is, the opacity and randomness of the generator are part of the method, and the Randomness Location Principle derived below loses its basis.
The reply is that the two individuations come apart only under specifiable conditions, and that where they do not come apart, the choice between them makes no difference to any modal or reliabilist verdict. For a reliabilist, what matters about a method is the proportion of true beliefs among the beliefs it produces, and for the modal conditions, it is whether the method yields false beliefs in nearby worlds. Neither depends on how often the method produces a belief at all. Write A for the event that is accepted and G for the full internal state and history of the generator, including its randomness. Say that acceptance screens off generation, in Reichenbach’s sense [41], when
When Equation (3) holds, the truth ratio and the modal profile of the whole pipeline coincide with those of . For modal evaluation, the corresponding conclusion requires the screening-off relation to remain stable across the nearby worlds relevant to sensitivity and safety. Under that condition, individuating the method more finely by including the generator yields the same modal verdict, so acceptance individuation loses nothing. Three conditions jointly suffice for screening off, and each corresponds to a component of the architecture.
- 1.
- No bypass. Every retained object has passed C. If some outputs are retained without checking, the generator determines which beliefs are formed, and its reliability enters through .
- 2.
- No channel into the checker. The verdict of C depends only on O and on a trusted base that the generator cannot alter: the generator cannot add axioms, modify the kernel or its libraries, or exploit an implementation defect in C. A search process rewarded for passing a checker has an incentive to exploit any discrepancy between the checker and the rules it is meant to implement, so this condition must be secured, for instance by rechecking in an isolated environment and listing every axiom the proof depends on. Where it fails, the generator enters through T.
- 3.
- Independent specification. The statement is fixed, or audited, independently of the generator. If the generator writes or translates the statement it then proves, it can make acceptance easy by making the statement weak, and it enters through S.
The whole-pipeline individuation is therefore the general case, and acceptance individuation is what it reduces to when conditions 1–3 hold. This reduction offers a principled, conditional response to the generality problem for the mathematical cases at issue: the relevant method type is the narrowest one whose truth ratio is invariant under variation in the variables that acceptance screens off. Equally important, the conditions say exactly where generation re-enters when screening fails: through when there is a bypass, through T when there is a channel into the checker, and through S when the specification is not independent. These are the three components on which Section 7 concentrates. The objection from the whole pipeline does not refute the framework. It marks out the boundary conditions under which the framework’s simplest verdicts hold.
5. Historical Evidence: The Migration of Epistemic Properties
This section presents the evidence for the Locus Thesis. In each episode, a property that PLP treats as a single attribute of a proof moves from one component of an architecture to another while the proposition stays fixed. Table 2 summarizes the cases.
5.1. The Generator–Checker Separation Before Machine Learning
The separation that matters most for AI mathematics was designed long before machines could exploit it at scale. The first heuristic theorem prover, the Logic Theorist of Newell, Shaw and Simon, found derivations for propositions in the second chapter of Principia Mathematica by heuristic search, sometimes by routes different from Whitehead and Russell’s [42,43]. The epistemically decisive step came with de Bruijn’s Automath and with the LCF (“Logic for Computable Functions”) tradition of proof assistants. De Bruijn required proof objects to be checkable by a program small enough for a skeptic to verify. In LCF-style systems, only a small trusted kernel can construct values of the type of theorems, so fallible higher-level search procedures cannot manufacture a false theorem [44,45,46]. In terms of , discovery can be assigned to arbitrary and possibly unreliable processes, while verification is assigned to a component whose reliability can be assessed independently. This is the engineering form of acceptance individuation, and every later architecture in this section inherits it.
5.2. Probability Inside the Argument: From Existence to Construction
Erdős’s 1947 lower bound for Ramsey numbers placed a probability space on graphs and showed that the probability of a bad coloring is less than one [47]. Nothing in the conclusion is probable; probability is a counting device inside a deductive argument [48]. The Lovász Local Lemma extends the method: if events each have probability at most p and each is mutually independent of all but at most d others, then implies that with positive probability none occurs. The classical proof establishes existence without identifying a witness.
Moser and Tardos showed that, in the variable model, repeatedly resampling the variables of any bad event that currently holds terminates quickly in expectation and returns a witness [49]. The proposition and its deductive status are unchanged; what changes is where randomness sits. The algorithm never returns an incorrect witness, and only its running time is random. Randomness has migrated from O to a generator whose output is checked deterministically.
5.3. One Predicate, Four Architectures: Primality
Primality is the cleanest natural experiment available, because the predicate is fixed while the other components vary.
A Miller–Rabin round writes with d odd and looks for a base a with and for . Such an a is a deterministic, cheaply checkable certificate of compositeness. Because at most a quarter of the bases are “strong liars” for an odd composite [50,51], k failures to find a witness bound the probability of a false “prime” verdict by . One procedure therefore has two outputs with different architectures. “Composite” comes with an object any reader can check. “Probably prime” comes with no object, only an error bound whose validity depends on the randomness of the bases.
Pratt proved that every prime has a succinct certificate: a base a with and for each prime q dividing , together with recursive certificates for those q [52]. Elliptic-curve primality proving (ECPP) replaces the multiplicative group with elliptic-curve groups and produces certificates that are practical for very large numbers [53]. The search for such certificates may be heuristic or randomized; their checking is deterministic. The AKS algorithm, finally, decides primality deterministically in polynomial time; decision and justification coincide in one procedure whose correctness rests on a published proof and on the implementation [54].
PLP must say of each of these four that “the proof” is or is not transferable, reliable or a priori, and the architectures show why the question is ill-posed: randomness sits in C for one Miller–Rabin output and nowhere for the other, in discovery only for Pratt and ECPP, and nowhere in AKS, whose trusted base instead contains the implementation of a complex algorithm.
5.4. The Four-Color Theorem: Surveyability Relocated Three Times
Appel and Haken’s argument had two parts [7,8]. A discharging argument, checkable by mathematicians in the ordinary way, showed that every minimal counterexample must contain a configuration from a finite unavoidable set. Computer programs then established the reducibility of each configuration. The computation discharged finitely many deductively sufficient obligations; it was not inductive extrapolation from a sample.
In 1977, surveyability was present for the global reduction (an informal O checked by humans) and absent for the reducibility checks (O the case data, C the programs, T code and hardware). Robertson, Sanders, Seymour and Thomas then gave a new proof with independent programs [55], an addition to R. Gonthier’s formalization in the Coq proof assistant relocated surveyability a third time [56]. It used proof by reflection, a technique in which one proves inside the proof assistant that a computable checking function is correct and then lets the kernel evaluate that function on the cases. The case checks remained humanly unsurveyable, but the trusted base shrank to the kernel and its execution environment, and the restatement of the theorem in terms of hypermaps shifted part of the residual doubt onto S.
The sequence records changes in which component carried which part of the doubt; at no stage did a non-proof turn into a proof.
5.5. Kepler and Flyspeck: When Refereeing Presupposes PLP
Hales announced a proof of the Kepler conjecture in 1998, combining conventional arguments with thousands of nonlinear inequalities verified by computer. After years of review, the referees of the Annals of Mathematics reported that they were “99% certain” of its correctness but could not certify every detail; the principal paper appeared in 2005 [57]. Peer review, as then organized, assumed that a referee could judge the proof as a whole; no referee could examine every component, and no procedure existed for composing component-level judgments.
Hales’s response, the Flyspeck project, was a re-architecture: definitions were redesigned for formal use and computational obligations recast so that the HOL Light and Isabelle proof assistants could certify them; the final account appeared in 2017 [58]. Trust moved from bespoke code and referee judgment into kernels, explicit definitions and verified computations, a migration that MacKenzie’s history of mechanized proof shows to be embedded in engineering cultures and negotiated standards [59]. The trusted base remained non-empty, and one can formally prove the wrong statement. Flyspeck relocated warrant; it did not abolish the need for it.
5.6. SAT Certificates: The Justificatory Object Becomes Unreadable
The Boolean satisfiability (SAT) problem asks whether a propositional formula, usually written in conjunctive normal form (CNF), has a satisfying assignment. Heule, Kullmann and Marek settled the Boolean Pythagorean Triples problem by encoding it as a SAT instance: the integers can be two-colored with no monochromatic Pythagorean triple, and cannot [60]. The unsatisfiability was certified by a refutation in DRAT format (a standard format that records each clause the solver learned so that an independent checker can re-derive the empty clause), close to 200 terabytes in size. Heule’s later determination of the fifth Schur number produced a certificate of the same character [61]. Here, O is beyond human reading, while C is small enough to be formally verified in Coq and independently in the ACL2 theorem prover [62]. The chain is
where is the propositional encoding and the refutation, and doubt concentrates on the first arrow. Later work accordingly verified the relation between the encoding and the combinatorial problem formally [63]. The episode is the clearest historical demonstration that, once O is unreadable and C verified, the live epistemic question becomes the specification.
5.7. Learning-Guided Generation
From about 2017, machine learning entered the generation side of the architecture: learned premise selection, tactic prediction and language-model proof search proposed steps that a proof assistant checked [64]. FunSearch coupled a language model that writes programs to an automated evaluator and found, among other things, a cap set in of size 512, larger than any previously known [65]. The finite claim is checkable by enumeration, and the model serves only as a proposal mechanism. An asymptotic consequence required a separate mathematical argument, so evaluator acceptance must be typed: a score may certify a finite object, suggest a pattern or support a theorem, and these are different achievements. AlphaProof searched for proofs in Lean and solved three of the six problems at the 2024 International Mathematical Olympiad; with AlphaGeometry 2, the combined system reached the silver-medal threshold, but the problems were translated into Lean by hand, and some searches ran for days [66,67]. In 2026, the First Proof project tested several AI systems on 10 unpublished research problems contributed by mathematicians and published problems, solutions, referee evaluations and system logs [68], and an agentic “AI co-mathematician” workbench combined ideation, literature search, computation, theorem proving and records of failed hypotheses [69].
These systems share the architecture that de Bruijn and LCF made possible. What changes, relative to every earlier episode, is the ratio between generation and checking: a learned generator can be run thousands of times, and human work concentrates on S (translating problems, writing evaluators) and on (deciding which outputs to retain and report).
The scale of the justificatory object, and the ratio between generation and checking, can be read off the published record. Table 3 collects figures reported in the primary sources for the episodes discussed. The figures show the machine-handled parts of the architecture growing by orders of magnitude while the part a human reads stays small or disappears, and in the most recent row, the quantity that has grown is the number of candidates, which is to say .
Table 2.
Migration of epistemic properties across components of in the historical cases.
Table 3.
Reported scale of selected episodes. Figures are those given in the cited sources.
6. Dissolving the Stalemates
6.1. Surveyability: Tymoczko and His Critics Described Different Components
Return to 1977 with the vocabulary of Section 3. Tymoczko’s claim that the warrant for the Four-Color Theorem contained an empirical element is true of T: that the computers executed the programs as written is an empirical proposition, and acceptance of the theorem depended on it [9]. Detlefsen and Luker’s claim that this was no novelty is true of O and C considered as types: ordinary proofs also depend on the correct execution of long calculations, and the reducibility checks were deductive obligations of an ordinary kind [30]. Teller’s claim that surveyability constrains coming to know, and not proofhood, is a claim about : it concerns which agent performs verification [29]. Burge’s claim that reliance on a properly functioning computational source need not make knowledge a posteriori concerns the entitlement attached to one component of T on the assumption that it functions properly [31].
These claims need not contradict one another when their different targets are distinguished. This does not remove substantive disagreements about mathematical knowledge or trust; it makes their points of contact more precise. Once relocated, a central remaining disagreement concerns how large is, who is in a position to estimate it, and whether independent replication (the 1997 re-proof, the 2005 formalization) reduces it enough to make the bound of Equation (2) negligible. These questions have answers. Bassler’s distinction reappears as a special case: global surveyability is a property of the informal layer of O together with S, and local surveyability a property of the remaining layers of O [32].
6.2. Transferability and the Randomness Location Principle
The probabilistic-proof debate yields to the same treatment, and here relocation produces a substantive principle.
Easwaran’s worry is that the reader of a probabilistic proof must accept on testimony that the random choices were random and independent of the instance [11]. In architectural terms, this is a worry about T (the randomness source belongs to the trusted base) and about (whether the reported run was selected from many). Jackson’s reply, that a reader can rerun the procedure with her own randomness, is a claim about R: it adds a replication route whose randomness the reader controls [34]. Easwaran is right that the object handed to the reader does not by itself warrant the conclusion. Jackson is right that the architecture as a whole can be made reader-checkable. They disagree only because each predicates transferability of “the probabilistic proof”.
Hamami’s lottery argument retains force for acceptance procedures with relevant false-acceptance possibilities [12]; Section 4 explains the modal qualification. Consider, again, the four primality architectures. In Miller–Rabin’s “probably prime” output, acceptance is compatible with falsity, with a known bound; the belief may share the modal vulnerability of a lottery belief where the relevant false-acceptance worlds are nearby. In ECPP, randomness may be used to find a certificate, but a found certificate is checked deterministically; the probability that the search succeeds has no bearing on the warrant that a successful certificate confers. The same holds for Moser–Tardos witnesses and FunSearch constructions.
Principle 2
(Randomness Location). Randomness in a warrant architecture bears on the warrant for P only if it is located in the acceptance method , including the randomness source on which an error bound for C depends. Randomness confined to the generation of O (to discovery, search or construction) is epistemically inert for the warrant that a successful deterministic check confers on .
The principle is a consequence of acceptance individuation together with sensitivity and safety, and so it is not a stipulation. It divides the probabilistic-proof literature along a line that neither side drew. Fallis’s examples of epistemically excellent probabilistic arguments [10] include cases where randomness is in C, and the lottery objection applies to those. A large and growing class of computation-dependent mathematics, including every certificate-producing method and nearly all current AI-assisted mathematics with formal checking, places randomness only in generation, and the lottery objection does not touch it. The principle also clarifies Paseau’s thesis that non-deductive evidence can yield mathematical knowledge [35]: if he is right, he is right about architectures with , and for most of the cases that make the topic urgent today the question does not arise at the level of C.
The principle has a corollary that drives Section 7. Inertness holds, given that the retained object is checked. If some outputs are retained without a check, generation re-enters the method through , and the question becomes how often the generator produces outputs that look acceptable but are false. Generation is inert only where acceptance screens it off.
6.3. Formal and Informal Proof: Tanswell’s Overgeneration Problem
A third debate shows the same pattern. On the derivation-indicator view, an informal proof is correct because it indicates the existence of a formal derivation [71]. Tanswell objected that many distinct formal derivations correspond to any informal proof, so the view cannot say which formal object the informal proof depends on [72]. Under PLP, this is a serious difficulty, because a unique formal proof must bear the credentials of the informal one. In the architectural framework, it is expected. An informal proof underdetermines O. What a good informal proof fixes is S and the global structure of O, leaving local layers to be filled by any member of a class of formal objects, each of which, if checked, warrants . The credentials of the informal proof consist in the existence and accessibility of that class together with the fidelity of S, and derivationism can be reconstructed without a unique derivation. Andersen’s analysis of acceptable gaps in published proofs fits the same structure: a gap is acceptable when the intended audience can fill it, that is, when the omitted layers fall within the competence of the agents to whom assigns verification [73].
7. What AI Changes: Specification and Selection
7.1. The Inversion of Epistemic Economy
Tao asks the mathematical community to suppose that AI systems will soon perform a significant share of research-level tasks, and argues that the decisive question then concerns the goals and values of mathematics [74]. Avigad urges mathematicians to understand and shape these technologies themselves [75,76]. Commelin, Jamnik, Ochigame, Taelman and Venkatesh warn that the capabilities of new technology risk dictating what counts as mathematically significant [77]. Pantsar proposes that AI-generated proofs be subjected to the same communal scrutiny as human ones [21]. The Leiden Declaration on Artificial Intelligence and Mathematics, endorsed by the International Mathematical Union, asks for disclosure of tool use, human responsibility for correctness and attribution, and rigorous standards for AI-assisted results [20,78].
The architectural framework locates these concerns. In every pre-2020 episode of Section 5, generation was scarce and expensive: a human or a dedicated program produced one justificatory object, or a few, for a problem chosen in advance. Under that condition, is degenerate, and S is written by people who understand the problem. When generation becomes cheap relative to checking, both conditions fail. Candidates appear in large numbers; statements are written, translated or chosen at scale; and the agents who check a formal object may differ from those who can judge whether its statement expresses anything significant.
7.2. Specification
Formal verification establishes . Whether is P is not something a kernel can decide, because P is not a formal object. The point is familiar from Flyspeck and from the SAT encodings, where the fidelity of the encoding required separate verification [58,63]. It becomes pressing when statements are produced by the pipeline that produces the proofs. AlphaProof’s Olympiad results relied on human translation of the problems into Lean, which kept S out of the generator’s reach [67]. When translation is automated, a system can produce a correct proof of a subtly different statement: a missing hypothesis, a degenerate case, and a library definition whose meaning differs from the informal one. Equation (2) shows that driving and towards zero leaves the bound at . Specification audit is the irreducible human contribution to formally verified AI mathematics, and it requires domain expertise, which proof-assistant expertise does not replace.
7.3. Selection and Release-Level Warrant
Suppose a process generates candidate results for N problems and a filter F decides which to retain. Let be the proportion of candidates that are correct, the power of the filter, and its false-retention rate. By Bayes’s theorem, the proportion of retained results that are correct is
This is the structure Ioannidis used to show that, in fields with a low prior probability of true hypotheses and weak filters, a large share of published positive findings can be false [79]. Two features matter for mathematics. For hard problems, is low, so even a small f can yield modest precision, and when F is a deterministic kernel applied to a faithfully specified statement, f is zero, and precision is one whatever and t are. This is the Randomness Location Principle applied to a population: generation affects and t, and acceptance fixes f.
Equation (5) has a consequence that makes release-level warrant partly measurable from public data. Let be the retention rate, the proportion of attempts that end in a retained result, which a laboratory can report. Since ,
The bound needs neither nor t, which are unobservable for open problems. It needs only the retention rate, which is public, and the false-retention rate, which can be estimated by intervention (below). The lower the retention rate, the more a given false-retention rate costs: a filter that wrongly retains one false candidate in a hundred is harmless when most attempts succeed and damaging when few do.
For a retained result that has not been checked by a deterministic procedure, the warrant available to a reader who has not checked it herself is, absent manuscript-specific evidence, bounded by the precision of the filter that retained it. I call this its release-level warrant. It is a property of , not of the manuscript, and it is invisible under PLP, which asks only whether this proof is correct.
Two consequences follow. First, a formally verified subset of a release certifies only itself. It is evidence about the unverified remainder only if the subset was chosen independently of the probability of correctness. If formalization was attempted first on results that were easier to formalize, or if items that resisted formalization were left in the unverified group, then the success of the verified subset overestimates the precision of the rest. Second, f can be estimated by intervention, and this is the right way to estimate it. The causal question behind release-level warrant is counterfactual: would this result have been retained had it been false? Interventionist accounts of causal explanation tell us to answer such questions by manipulating the putative cause while holding the rest of the system fixed [80]. Applied here, this suggests decoy injection: introduce into the generation pipeline a known number of statements that are false (with subtly broken hypotheses, for instance) or already settled, mixed blindly with genuine problems, and measure how often the pipeline retains a purported proof of a false decoy. The observed rate estimates f directly, which no amount of inspection of the retained successes can do. Combined with an independent expert audit of a random sample of unverified retained items, it yields an estimate of release-level warrant with a stated uncertainty.
7.4. The October 2026 Release as a Test Case
The release described in the Introduction fits this analysis closely. Its public description reports approximately 4000 problems posed, 722 manuscripts in 372 families, results “at different stages of verification”, a Lean library in which “many, but not all” of the manuscripts have been formalized, and an explicit warning that some unformalized results “could have issues” [1]. I make no claim about the correctness of any individual result, and none is needed. The release is philosophically instructive because it makes the components of W visibly separate.
For the formalized results, C and T are those of Lean and its library, which are well understood; the decisive question is S, namely, whether each Lean statement expresses the mathematical claim stated in the manuscript, and answering it requires experts in the relevant fields. For the unformalized results, the warrant available to a reader who has not checked a manuscript is release-level warrant, and its size depends on quantities the public description does not fix: the rule that maps roughly 4000 posed problems to 722 manuscripts, and the false-retention rate of whatever internal checks were applied. The replication set R is constrained in a specific way. Because the model is unreleased, independent parties can recheck outputs but cannot regenerate them, so replication can bear on C and not on the generating process. Finally, assigns problem formation, specification audit and stabilization largely to humans outside the producing organization, which is the role the IAS Advisory Group was constituted to play [2].
The reported figures allow a first quantitative reading. If the 372 result families are taken as a proxy for problems on which a result was retained, the problem-level retention rate is about . Equation (6) then fixes how much release-level warrant depends on the unknown false-retention rate of whatever internal checks governed the unformalized items. Table 4 gives the lower bound for illustrative values of f. These values are hypothetical; no false-retention rate has been reported. The table shows what would need to be measured, not what is the case. It also shows why the figure matters: at the observed retention rate, the difference between a false-retention rate of one in a hundred and one in twenty is the difference between a lower bound of about 0.89 and a lower bound of about 0.46.
Table 4.
Lower bound on release-level precision, Equation (6), at the reported retention rate , for illustrative (hypothetical) false-retention rates f.
For the formalized portion, the bound is irrelevant, since a kernel applied to a faithful specification has ; this is why the proportion formalized (about 22 percent of manuscripts, by press reports [70]) matters so much, and why, as argued above for verified subsets, it says nothing by itself about the remaining 78 percent.
The questions that matter for such a release cannot be posed under PLP. “Are these proofs correct?” has 722 answers and no general one. “What is the warrant architecture of this release, and which of its components have been audited, by whom and how?” has a definite answer, and it is the question that editors, readers and advisory bodies need answered.
7.5. The Controversy of October 2026 as a Test of the Locus Thesis
The public reaction to the release provides an unplanned test of the Locus Thesis. If PLP were adequate, the controversy should have turned on a single question, whether the results are correct, with positions ranged along one axis of confidence. It did not. Each major position addressed a different component of the architecture, and positions that look opposed are in several cases compatible once relocated.
Selection. The guidelines published by the Advisory Group on 29 September ask laboratories that release batches of results to publish a companion document stating how the problems were selected and “how many other problems of comparable difficulty the models tried and failed to solve”, together with, for each result, the model, the prompts, a summarized chain of thought, the time taken and the estimated computing cost [3]. These requests concern and nothing else. The release disclosed a denominator (roughly 4000 problems posed) and an average compute figure [1]. According to press reports, it did not publish prompts; it provided reasoning summaries for 10 results, and OpenAI described almost all results as obtained from a single prompt to a single agent [70]. The Advisory Group’s own statement of 6 October left the judgment of compliance to the community and described the release as “the beginning, not the completion” of the work of understanding it [81]. In the terms of Section 7, a body of mathematicians independently demanded exactly the disclosures on which release-level warrant depends.
Specification and checking. Press reports put the number of manuscripts whose main result is formalized in Lean at 162 of 722, about 22 percent [70]; the repository itself says only “many, but not all” [1]. Sutherland’s demand that mathematicians “should ask for receipts” before treating one-shot claims as established [70] is a demand about the unformalized remainder, where release-level warrant is all a reader has. For the formalized portion, it is satisfied by the kernel, subject to the specification audit that no kernel performs.
Replication. Because the model is unreleased, outsiders can recheck outputs but cannot regenerate them. Calls to treat the results as unverified until the model is available [70] are claims about R, and by the Randomness Location Principle, they have force only for results that acceptance does not screen off.
Allocation: credit, novelty and stabilization. Buckmaster’s complaint that OpenAI had not done “due diligence” on whether results reproduced or completed existing work [4] concerns problem formation and stabilization in , the functions that assign credit and situate a result in the literature. It is independent of correctness: a fully verified result can fail it.
Explanation. Tao’s response was that a “Math 2.0” will need to “decenter the role of raw problem solving” in favor of exposition, community building and new directions [6]. The Advisory Group’s guidelines likewise require laboratories that release output no one yet understands to “take responsibility for ensuring that human understanding will follow” [3]. Both concern the explanation function, which, as Section 8 argues, no amount of verification supplies.
Governance. The statement of the Association for Human Mathematics, published on Tao’s blog on 7 October, declared that “releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power” and urged mathematicians “to discontinue their work with OpenAI” [5]. Its target is the allocation of authority over problem selection and over the infrastructure of mathematics, a political question that Objection 8 below separates from the epistemic one.
Access. Litt, finally, judged the release “great for mathematics” and saw no reason to keep the answers to mathematical questions secret [4]. His claim concerns the availability of O to the community, which is a precondition of every audit listed above.
These positions do not lie on a single scale of optimism. Litt can be right about access while Buckmaster is right about credit, Sutherland about the unformalized remainder, Tao and the Advisory Group about explanation, and the Association for Human Mathematics about governance. The controversy looks intractable only when it is read as a dispute over whether “the results” are good. Read component by component, it becomes a list of audits, each of which has an owner and an answer.
7.6. Where Generator Opacity Matters
The generators in these systems are epistemically opaque in Humphreys’s sense: no human agent knows all the epistemically relevant elements of the process when it runs [22]. Creel distinguishes functional, structural and run transparency in complex computational systems [23], and Durán and Formanek argue that reliability established through verification, robustness analysis, a track record and expert knowledge can justify trust in computational results despite essential opacity [24]. Friedman describes the anxiety this opacity produces in mathematical practice, where the mathematician’s role shifts towards checking and steering outputs she did not produce [19]. The present framework sharpens these results for mathematics. By acceptance individuation and the Randomness Location Principle, opacity of generation is inert for the warrant of a deterministically checked result. It matters in exactly two places: in , where the reliability of the generator determines the precision of retention filters for unchecked outputs, and in explanation, where an opaque generator provides no account of why a result holds. The first is addressed by decoy injection and by the structural tests of reliability, robustness under problem-preserving perturbation, counterfactual sensitivity and transfer to held-out problem families proposed in earlier work on xenoepistemics [16]. The second is the subject of the next section.
8. Warrant Is Not Explanation: Xenoepistemics and Co-Mathematics
Discussions of AI in mathematics frequently call for “explainability”. The framework shows why the call is ambiguous. Two demands are commonly run together under the label “explainable AI”.
- 1.
- Generator assessment: Is the process that produced candidates reliable, robust, counterfactually sensitive and capable of transfer? This bears on warrant only through , that is, only where acceptance does not screen generation off. The xenoepistemic tests belong here. They evaluate a knowledge producer whose operations need not be humanly intelligible, which is the defining situation of xenoepistemics [16].
- 2.
- Explanatory provision: Does the community possess an account of why the result holds that supports compression, navigation and reuse? This is the explanation function in . It is a requirement on mathematics as a shared practice, and it is a requirement on whoever is assigned that function, human or machine.
The two come apart in both directions. A deterministically checked result from an opaque generator can be fully warranted and wholly unexplained; the 200-terabyte refutation is the limiting case. Conversely, a lucid explanation can accompany a false claim. Keeping the demands apart yields a division of labor. Xenoepistemics concerns when non-human processes can be trusted as sources of mathematical warrant without being understood. The demands for navigable, abstraction-respecting explanations concern what I have called co-mathematics: the allocation of the explanation function in a hybrid practice, where machines may contribute but where the test of success lies in human uptake.
That test can itself be made explicit. An explanation, whoever produces it, succeeds to the degree that it supports uptake: readers who have studied it can solve related problems, identify which hypotheses are used where, predict the effect of weakening a hypothesis, and transfer the method to a neighboring question. Such capacities, which include navigating an argument at the level of detail one needs and grasping its useful levels of abstraction, are the ones that Thurston took to be the point of proof [25], that Rav located in proofs as bearers of mathematical knowledge [26], and that Friedman and Kish Bar-On ask us to rethink as mathematical intuition changes under AI [82]. The Liquid Tensor Experiment, in which a formalization organized by explicit abstraction boundaries changed how the mathematics was understood, shows that formalization and explanation can reinforce one another when the explanation function is deliberately assigned [27]. Uptake tests are tests of explanations and of their human users. Xenoepistemic tests are tests of generators. A mature hybrid practice needs both, and it should not mistake success on one for success on the other.
9. Layered Auditability
If surveyability by a single competent reader is no longer a reasonable requirement, something must replace it, on pain of having no criterion at all. I propose the following.
Layered auditability. A claim P is acceptable on the basis of when (i) coverage: every component of W has been audited by at least one qualified agent independent of the generator; (ii) typed interfaces: each audit states its output and its assumptions in a form the audit of the adjacent component can take as input; (iii) composition: the component audits jointly license a global error assessment by rules such as Equation (2); (iv) independence: wherever the claim’s importance warrants it, R contains at least one route whose failure modes are materially independent of the original; and (v) disclosure: is reported whenever any retained component has not been deterministically checked.
The criterion is fallibilist. De Toffoli argues that the primary role of mathematical justification is to enable a community of fallible agents to detect and correct errors [83]. Layered auditability specifies the structure that makes such detection possible when no individual can inspect everything: coverage ensures that every place an error could hide has an auditor, and typed interfaces ensure that an error detected in one component is communicated to the audits that depend on it. On this reading, De Toffoli’s fallibilism and the Locus Thesis support each other. Fallibilism says what justification is for, and the architecture says where in a hybrid proof the fallibilist’s error-detecting capacities must be located.
The criterion is social without being merely sociological. Habgood-Coote and Tanswell argue that the Classification of Finite Simple Groups is known only by a group whose members have partial access [18], and Hardwig showed that rational belief often depends on others’ expertise [84]. Layered auditability states when such group knowledge is well formed: when the partial accesses are typed and composed. It is continuous with practice as well. Avigad’s account of reliability through modularity and redundancy [36] and Andersen’s account of acceptable gaps [73] describe special cases.
The criterion has an operational form. Table 5 gives a minimal warrant chart that authors of computationally or AI-assisted results could submit with a manuscript, so that editors can assign referees to components. A domain referee audits S, significance and explanation; a formalization or computation referee audits O, C, T, and R; for releases at scale, a statistical audit of is added. Acceptance labels can then be precise (formally certified, certificate-checked, independently reproduced, expert-adjudicated, explanatorily reconstructed) and cumulative only where evidence supports each.
Table 5.
Minimal warrant chart for a computationally or AI-assisted mathematical claim.
10. Hard Cases: Positions on Contested Questions
The following five cases show how the proposed distinctions may guide contested decisions. They are illustrative applications, not a comprehensive theory of authorship, editorial governance, institutional responsibility or AI co-mathematics. Each topic raises further normative questions beyond the central account of mathematical warrant.
- Should journals publish unformalized AI-generated results?
Yes, as candidate results with an explicit label, and no, as theorems, until manuscript-level checking has replaced release-level warrant. The position follows from Equation (5): a reader of a journal is entitled to assume that publication certifies a manuscript, and a release-level filter of unknown precision does not license that assumption. A label such as “expert-adjudicated” should therefore be reserved for results that have actually been adjudicated.
- Who decides when the formalizer and the domain expert disagree about S?
The domain expert has the final word on whether expresses P, and the formalizer has the final word on whether C accepted a term for . Disagreements are typically about which component is at fault, and the warrant chart makes that question explicit. A kernel’s acceptance can never settle a dispute about S.
- Can an AI system be an author?
Authorship should track responsibility for the functions that require it: problem formation, specification audit, response to criticism and stabilization. A system that generated a proof performed the discovery and justification functions and should be credited for them in a contribution statement, as a computer-assisted proof credits its programs. It cannot answer for S or respond to a referee, so it cannot bear authorship in the sense that the Leiden Declaration and editorial practice require [20]. The allocation lets credit and responsibility be assigned separately, which the binary “author or tool” does not.
- Are results from unreleased models acceptable?
For checked results, yes, by the Randomness Location Principle, provided that the screening-off conditions of Section 4 are secured, for instance, by rechecking in an independent environment: the warrant of such a result does not depend on access to the generator. For unchecked results, release-level warrant cannot be estimated by outsiders without regeneration or decoy injection, so such results should carry no more than the label “reported”.
- Does layered auditability make mathematical knowledge irreducibly testimonial?
It makes it irreducibly distributed, which is a different claim. Each component audit is first-hand for the auditor, and composition is checkable by anyone who can read the chart. What no individual has is first-hand access to every component, and Burge’s analysis of entitlement shows that this lack need not demote the resulting knowledge to the a posteriori [31].
11. Objections and Replies
- Objection 1: The kernel suffices.
Once a statement has been faithfully formalized and a sound kernel has accepted a derivation, the mathematical question is settled; everything else is sociology. The formal turn is valuable because it separates objective derivability from variable judgments [85].
Reply. The objection is correct about formal proofhood, and the Locus Thesis concedes it: relative to a formal system and a correct kernel, acceptance establishes . However, its antecedent includes “faithfully formalized”, a clause that names a component, S, that no kernel checks. The objection presupposes the decomposition it was meant to resist. What it shows is that in formally verified mathematics the residual epistemic question lies wholly in S and T, which is a consequence of the Locus Thesis.
- Objection 2: Transferability is irreducibly global.
Easwaran’s notion is defined for a proof and a reader: the reader comes to know the conclusion from the proof without testimony. Decomposing the proof loses the property.
Reply. Consider Flyspeck. No reader comes to know the Kepler conjecture from the formal proof without relying on the kernel, the compiler and the hardware, so on the global reading Flyspeck is not transferable. Either Flyspeck confers less warrant than the partly unverified 2005 paper that the Annals referees could certify only to “99%” [57,58], which no one believes, or transferability must be read component-wise: the reader can check S, audit the informal layer of O, and rely on a component of T whose reliability has been independently established. The second reading keeps what was valuable in Easwaran’s notion, independence from the producer’s testimony, and drops the requirement that a single reader verify everything.
- Objection 3: Burgean entitlement makes the architecture idle.
If we are entitled a priori to rely on properly functioning computational sources as we rely on interlocutors [31], the decomposition adds nothing; the mathematician simply knows.
Reply. Entitlement of this kind is defeasible and conditioned on proper functioning, and it does not reach defeaters that arise from how a source’s outputs were selected. A generator that is reliable in general but whose outputs are filtered by an undisclosed rule can deliver a sample in which errors are overrepresented, and the entitlement to rely on the source does not address that. The objection is right for C and T under proper functioning; it says nothing about S or .
- Objection 4: Unformalized AI proofs are lottery propositions.
Extending Hamami [12], one might argue that accepting an unformalized result from a mass release on the strength of the release’s reliability is believing a lottery proposition, and so cannot yield knowledge.
Reply. The structures differ. A lottery proposition is believed on statistical grounds that are, by construction, insensitive to the particular outcome. An unformalized manuscript contains a specific argument that any qualified reader can check. Release-level warrant is the warrant available before such checking, and the right conclusion is that it does not suffice for knowledge when the precision of the release is unknown. That is compatible with knowledge being obtained by checking the manuscript, which turns release-level into manuscript-level warrant. The objection supports the recommendations of Section 7 and Section 10, and gives no reason for wholesale rejection.
- Objection 5: Nothing new.
Philosophers of mathematics discussed computer proofs and probabilistic proofs decades before generative AI. Calling distributed warrant an “architecture” renames a known phenomenon.
Reply. The objection is right that machine participation in proof is old, and Section 5 is designed to show it. Four things are new: the diagnosis of PLP as the common root of two stalemates, together with a prediction of which parts of each survive relocation; acceptance individuation as an answer to the generality problem for mathematical reliabilism, with explicit screening-off conditions under which it holds; the Randomness Location Principle derived from it, which divides the probabilistic-proof literature along a line neither side drew; and the identification of specification and selection as the components that cheap generation makes dominant, with an interventionist procedure for auditing the latter. The classical literature assumed scarce generation, and so treated as degenerate and S as unproblematic.
- Objection 6: Opacity defeats layering.
Learned generators are essentially opaque [22]. If one component cannot be inspected, layered auditability fails its own coverage condition.
Reply. Coverage requires that every component be audited; it does not require that every component be transparent. For checked results, the generator is no component of the acceptance method; by acceptance individuation, it is part of the history of O, and it is inert once C has accepted O. Where the generator does matter, in , it is audited behaviorally, by decoy injection and the xenoepistemic tests, as computational reliabilism audits simulations whose internals no one inspects [24].
- Objection 7: The framework is individualist in disguise.
Mathematical knowledge is irreducibly social [18]; decomposing warrant into components that individuals audit reinstates the lone knower on a smaller scale.
Reply. The framework includes the social components explicitly: records the division of epistemic labor, R records communal replication, and the composition condition of layered auditability is satisfied by a group and by no single member. What it adds to social accounts is a structural condition on when group knowledge is well formed. A group whose partial accesses do not compose, because no one has audited S or because is undisclosed, does not know P merely because each member knows something.
- Objection 8: Human governance is an anthropocentric residue.
If generated results can be warranted without human understanding, as the framework allows, insisting that humans govern mathematical institutions is arbitrary.
Reply. The epistemic and the political questions come apart. Xenoepistemics rejects human understanding as a necessary condition of warrant for checked results [16]. It does not follow that systems which cannot bear responsibility should hold authority over problem formation, specification audit, attribution and stabilization, the functions that assigns to humans in every case examined here. The reason is accountability and the public purposes of mathematics, which the Leiden Declaration, the advisory structures of 2026 and the controversy of October 2026 make explicit [3,5,20]. No claim of cognitive superiority is involved.
12. Conclusions
The philosophy of computer proof began with a question about one object: is this unsurveyable proof a proof? The philosophy of probabilistic proof began with another: is this probabilistic proof transferable? Both questions presuppose proof-level predication, and both debates stalled because each side was right about a different component of one architecture. The Locus Thesis replaces the presupposition with a decomposition as well as composition rules, and acceptance individuation grounds it in the theory of knowledge: the method that confers warrant is the acceptance procedure, and generation belongs to it only where acceptance fails to screen it off. Applied to history, the thesis shows the disputed properties migrating between components from 1947 to 2026. Applied to the stalemates, it separates their verbal from their substantive parts and yields the Randomness Location Principle. Applied to the present, it turns the controversy over OpenAI’s 722 manuscripts into a set of component audits, and it identifies what cheap generation changes: the decisive loci of doubt become the specification, which no kernel checks, and the selection profile, which no single manuscript reveals and which can be audited by intervention.
The result is a proposal for an epistemology of mathematics suited to computational environments. Knowledge is produced by architectures in which computation is constitutive. Some components are opaque to every human participant, and the productive question about any result is which component carries which part of the warrant and who has audited it. Such an epistemology admits non-human processes as sources of warrant and treats explanation as a separate achievement that someone must be assigned to provide. A release of hundreds of machine-produced manuscripts can enlarge the stock of verified statements without enlarging understanding. Whether it enlarges understanding depends on whether the community audits specifications, demands selection records and assigns explanation to agents who can carry it out. Replacing surveyability with layered auditability makes these tasks explicit, and they fall to mathematicians and philosophers working together.
Funding
This research received no external funding.
Institutional Review Board Statement
Not applicable.
Informed Consent Statement
Not applicable.
Data Availability Statement
No new data were created or analyzed in this study. Data sharing is not applicable to this article.
Acknowledgments
As a non-native English speaker, the author used an AI-assisted language tool for English-language proofreading, stylistic correction and bibliographic checking. The tool was not used to generate the scientific claims or arguments. The author reviewed all changes and assumes full responsibility for the final content.
Conflicts of Interest
The author declares no conflicts of interest.
Abbreviations
The following abbreviations are used in this manuscript:
| AI | Artificial intelligence |
| AKS | Agrawal–Kayal–Saxena primality test |
| CNF | Conjunctive normal form |
| DRAT | Deletion Resolution Asymmetric Tautology (proof format for SAT solvers) |
| ECPP | Elliptic-curve primality proving |
| LCF | Logic for Computable Functions (proof-assistant architecture) |
| LLL | Lovász Local Lemma |
| PLP | Proof-level predication |
| SAT | Boolean satisfiability |
References
- OpenAI. Openai/math: Mathematical Manuscripts and Supporting Proof Artifacts Produced by an Internal OpenAI Model. 6 October 2026. Available online: https://github.com/openai/math (accessed on 7 October 2026).
- OpenAI. Advisory Group on Mathematics and Artificial Intelligence. 21 September 2026. Available online: https://openai.com/index/advisory-group-on-mathematics-and-ai/ (accessed on 7 October 2026).
- Advisory Group on Mathematics and Artificial Intelligence. Responsible Release of AI-Generated Mathematics. 29 September 2026. Available online: https://agmai.org/general-sep29/ (accessed on 8 October 2026).
- Fortune. OpenAI Publishes Solutions to More than 370 Outstanding Math Challenges. Math May Never Be the Same. 7 October 2026. Available online: https://fortune.com/2026/10/07/openai-math-controversy-solutions-370-outstanding-challenges-published-criticisms-celebration/ (accessed on 8 October 2026).
- Association for Human Mathematics Communications Working Group. AHM Statement on OpenAI’s October 6 Release of Mathematical Documents. Guest Post on T. Tao, What’s New. 7 October 2026. Available online: https://terrytao.wordpress.com/2026/10/07/ahm-statement-on-openais-october-6-release-of-mathematical-documents/ (accessed on 8 October 2026).
- Tao, T. Post on Mathstodon. 6 October 2026. Available online: https://mathstodon.xyz/@tao/117395269325940185 (accessed on 8 October 2026).
- Appel, K.; Haken, W. Every Planar Map Is Four Colorable. Part I: Discharging. Ill. J. Math. 1977, 21, 429–490. [Google Scholar] [CrossRef] [Scilit]
- Appel, K.; Haken, W.; Koch, J. Every Planar Map Is Four Colorable. Part II: Reducibility. Ill. J. Math. 1977, 21, 491–567. [Google Scholar] [CrossRef] [Scilit]
- Tymoczko, T. The Four-Color Problem and Its Philosophical Significance. J. Philos. 1979, 76, 57–83. [Google Scholar] [CrossRef] [Scilit]
- Fallis, D. The Epistemic Status of Probabilistic Proof. J. Philos. 1997, 94, 165–186. [Google Scholar] [CrossRef] [Scilit]
- Easwaran, K. Probabilistic Proofs and Transferability. Philos. Math. 2009, 17, 341–362. [Google Scholar] [CrossRef] [Scilit]
- Hamami, Y. Probabilistic Proofs, Lottery Propositions, and Mathematical Knowledge. Philos. Q. 2022, 72, 77–89. [Google Scholar] [CrossRef] [Scilit]
- Conee, E.; Feldman, R. The Generality Problem for Reliabilism. Philos. Stud. 1998, 89, 1–29. [Google Scholar] [CrossRef] [Scilit]
- Humphreys, P. Extending Ourselves: Computational Science, Empiricism, and Scientific Method; Oxford University Press: New York, NY, USA, 2004. [Google Scholar]
- Vallverdú i Segura, J. Computational Epistemology and e-Science: A New Way of Thinking. Minds Mach. 2009, 19, 557–567. [Google Scholar] [CrossRef] [Scilit]
- Vallverdú, J. Xenoepistemics. Philosophies 2026, 11, 57. [Google Scholar] [CrossRef] [Scilit]
- Hutchins, E. Cognition in the Wild; MIT Press: Cambridge, MA, USA, 1995. [Google Scholar]
- Habgood-Coote, J.; Tanswell, F.S. Group Knowledge and Mathematical Collaboration: A Philosophical Examination of the Classification of Finite Simple Groups. Episteme 2023, 20, 281–307. [Google Scholar] [CrossRef] [Scilit]
- Friedman, M. Mathematical Anxiety in the Age of AI: Opacity, Role Inversion, and the Changing Place of the Mathematician. Comput. Hum. Behav. Artif. Hum. 2026, 9, 100386. [Google Scholar] [CrossRef] [Scilit]
- Leiden Declaration Working Group. Leiden Declaration on Artificial Intelligence and Mathematics; Zenodo: Geneva, Switzerland, 2026. [Google Scholar] [CrossRef]
- Pantsar, M. Theorem Proving in Artificial Neural Networks: New Frontiers in Mathematical AI. Eur. J. Philos. Sci. 2024, 14, 4. [Google Scholar] [CrossRef] [Scilit]
- Humphreys, P. The Philosophical Novelty of Computer Simulation Methods. Synthese 2009, 169, 615–626. [Google Scholar] [CrossRef] [Scilit]
- Creel, K.A. Transparency in Complex Computational Systems. Philos. Sci. 2020, 87, 568–589. [Google Scholar] [CrossRef] [Scilit]
- Durán, J.M.; Formanek, N. Grounds for Trust: Essential Epistemic Opacity and Computational Reliabilism. Minds Mach. 2018, 28, 645–666. [Google Scholar] [CrossRef] [Scilit]
- Thurston, W.P. On Proof and Progress in Mathematics. Bull. Am. Math. Soc. 1994, 30, 161–177. [Google Scholar] [CrossRef] [Scilit]
- Rav, Y. Why Do We Prove Theorems? Philos. Math. 1999, 7, 5–41. [Google Scholar] [CrossRef] [Scilit]
- Commelin, J.; Topaz, A. Abstraction Boundaries and Spec Driven Development in Pure Mathematics. arXiv 2023, arXiv:2309.14870. [Google Scholar] [CrossRef] [Scilit]
- Wittgenstein, L. Remarks on the Foundations of Mathematics, rev. ed.; von Wright, G.H., Rhees, R., Anscombe, G.E.M., Eds.; Anscombe, G.E.M., Translator; Blackwell: Oxford, UK, 1978. [Google Scholar]
- Teller, P. Computer Proof. J. Philos. 1980, 77, 797–803. [Google Scholar] [CrossRef] [Scilit]
- Detlefsen, M.; Luker, M. The Four-Color Theorem and Mathematical Proof. J. Philos. 1980, 77, 803–820. [Google Scholar] [CrossRef] [Scilit]
- Burge, T. Computer Proof, Apriori Knowledge, and Other Minds. Philos. Perspect. 1998, 12, 1–37. [Google Scholar] [CrossRef] [Scilit]
- Bassler, O.B. The Surveyability of Mathematical Proof: A Historical Perspective. Synthese 2006, 148, 99–133. [Google Scholar] [CrossRef] [Scilit]
- Fallis, D. The Reliability of Randomized Algorithms. Br. J. Philos. Sci. 2000, 51, 255–271. [Google Scholar] [CrossRef] [Scilit]
- Jackson, J.C. Randomized Arguments Are Transferable. Philos. Math. 2009, 17, 363–368. [Google Scholar] [CrossRef] [Scilit]
- Paseau, A. Knowledge of Mathematics without Proof. Br. J. Philos. Sci. 2015, 66, 775–799. [Google Scholar] [CrossRef] [Scilit]
- Avigad, J. Reliability of Mathematical Inference. Synthese 2021, 198, 7377–7399. [Google Scholar] [CrossRef] [Scilit]
- Nozick, R. Philosophical Explanations; Harvard University Press: Cambridge, MA, USA, 1981. [Google Scholar]
- Sosa, E. How to Defeat Opposition to Moore. Philos. Perspect. 1999, 13, 141–153. [Google Scholar] [CrossRef] [Scilit]
- Williamson, T. Knowledge and Its Limits; Oxford University Press: Oxford, UK, 2000. [Google Scholar]
- Pritchard, D. Epistemic Luck; Oxford University Press: Oxford, UK, 2005. [Google Scholar]
- Reichenbach, H. The Direction of Time; University of California Press: Berkeley, CA, USA, 1956. [Google Scholar]
- Newell, A.; Simon, H.A. The Logic Theory Machine—A Complex Information Processing System. IRE Trans. Inf. Theory 1956, 2, 61–79. [Google Scholar] [CrossRef] [Scilit]
- Newell, A.; Shaw, J.C.; Simon, H.A. Empirical Explorations of the Logic Theory Machine: A Case Study in Heuristic. In Proceedings of the Western Joint Computer Conference; ACM: New York, NY, USA, 1957; pp. 218–230. [Google Scholar] [CrossRef] [Scilit]
- de Bruijn, N.G. The Mathematical Language AUTOMATH, Its Usage, and Some of Its Extensions. In Symposium on Automatic Demonstration; Lecture Notes in Mathematics, Vol. 125; Springer: Berlin/Heidelberg, Germany, 1970; pp. 29–61. [Google Scholar]
- Gordon, M.J.C.; Milner, R.; Wadsworth, C.P. Edinburgh LCF: A Mechanised Logic of Computation; Lecture Notes in Computer Science, Vol. 78; Springer: Berlin/Heidelberg, Germany, 1979. [Google Scholar]
- Barendregt, H.; Wiedijk, F. The Challenge of Computer Mathematics. Philos. Trans. R. Soc. A 2005, 363, 2351–2375. [Google Scholar] [CrossRef] [Scilit] [PubMed]
- Erdős, P. Some Remarks on the Theory of Graphs. Bull. Am. Math. Soc. 1947, 53, 292–294. [Google Scholar] [CrossRef] [Scilit]
- Alon, N.; Spencer, J.H. The Probabilistic Method, 4th ed.; Wiley: Hoboken, NJ, USA, 2016. [Google Scholar]
- Moser, R.A.; Tardos, G. A Constructive Proof of the General Lovász Local Lemma. J. ACM 2010, 57, 11. [Google Scholar] [CrossRef] [Scilit]
- Rabin, M.O. Probabilistic Algorithm for Testing Primality. J. Number Theory 1980, 12, 128–138. [Google Scholar] [CrossRef] [Scilit]
- Monier, L. Evaluation and Comparison of Two Efficient Probabilistic Primality Testing Algorithms. Theor. Comput. Sci. 1980, 12, 97–108. [Google Scholar] [CrossRef] [Scilit]
- Pratt, V.R. Every Prime Has a Succinct Certificate. SIAM J. Comput. 1975, 4, 214–220. [Google Scholar] [CrossRef] [Scilit]
- Atkin, A.O.L.; Morain, F. Elliptic Curves and Primality Proving. Math. Comp. 1993, 61, 29–68. [Google Scholar] [CrossRef]
- Agrawal, M.; Kayal, N.; Saxena, N. PRIMES Is in P. Ann. Math. 2004, 160, 781–793. [Google Scholar] [CrossRef] [Scilit]
- Robertson, N.; Sanders, D.P.; Seymour, P.D.; Thomas, R. The Four-Colour Theorem. J. Comb. Theory Ser. B 1997, 70, 2–44. [Google Scholar] [CrossRef] [Scilit]
- Gonthier, G. Formal Proof—The Four-Color Theorem. Not. Am. Math. Soc. 2008, 55, 1382–1393. [Google Scholar]
- Hales, T.C. A Proof of the Kepler Conjecture. Ann. Math. 2005, 162, 1065–1185. [Google Scholar] [CrossRef] [Scilit]
- Hales, T.; Adams, M.; Bauer, G.; Dang, T.D.; Harrison, J.; Hoang, L.T.; Kaliszyk, C.; Magron, V.; McLaughlin, S.; Nguyen, T.T.; et al. A Formal Proof of the Kepler Conjecture. Forum Math. Pi 2017, 5, e2. [Google Scholar] [CrossRef] [Scilit]
- MacKenzie, D. Mechanizing Proof: Computing, Risk, and Trust; MIT Press: Cambridge, MA, USA, 2001. [Google Scholar]
- Heule, M.J.H.; Kullmann, O.; Marek, V.W. Solving and Verifying the Boolean Pythagorean Triples Problem via Cube-and-Conquer. In Theory and Applications of Satisfiability Testing—SAT 2016; Lecture Notes in Computer Science, Vol. 9710; Springer: Cham, Switzerland, 2016; pp. 228–245. [Google Scholar] [CrossRef] [Scilit]
- Heule, M.J.H. Schur Number Five. Proc. AAAI Conf. Artif. Intell. 2018, 32, 6598–6606. [Google Scholar] [CrossRef] [Scilit]
- Cruz-Filipe, L.; Heule, M.J.H.; Hunt, W.A., Jr.; Kaufmann, M.; Schneider-Kamp, P. Efficient Certified RAT Verification. In Automated Deduction—CADE 26; Lecture Notes in Computer Science, Vol. 10395; Springer: Cham, Switzerland, 2017; pp. 220–236. [Google Scholar] [CrossRef] [Scilit]
- Cruz-Filipe, L.; Schneider-Kamp, P. Formally Proving the Boolean Pythagorean Triples Conjecture. In LPAR-21: 21st International Conference on Logic for Programming, Artificial Intelligence and Reasoning; EPiC Series in Computing; EasyChair: Stockport, UK, 2017; Volume 46, pp. 509–522. [Google Scholar] [CrossRef] [Scilit]
- Polu, S.; Sutskever, I. Generative Language Modeling for Automated Theorem Proving. arXiv 2020, arXiv:2009.03393. [Google Scholar] [CrossRef] [Scilit]
- Romera-Paredes, B.; Barekatain, M.; Novikov, A.; Balog, M.; Kumar, M.P.; Dupont, E.; Ruiz, F.J.R.; Ellenberg, J.S.; Wang, P.; Fawzi, O.; et al. Mathematical Discoveries from Program Search with Large Language Models. Nature 2024, 625, 468–475. [Google Scholar] [CrossRef] [Scilit] [PubMed]
- Hubert, T.; Mehta, R.; Sartran, L.; Horváth, M.Z.; Žužić, G.; Wieser, E.; Huang, A.; Schrittwieser, J.; Schroecker, Y.; Masoom, H.; et al. Olympiad-Level Formal Mathematical Reasoning with Reinforcement Learning. Nature 2025, 651, 607–613. [Google Scholar] [CrossRef] [Scilit] [PubMed]
- Google DeepMind. AI Achieves Silver-Medal Standard Solving International Mathematical Olympiad Problems. 25 July 2024. Available online: https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/ (accessed on 7 October 2026).
- Abouzaid, M.; Srivastava, N.; Ward, R.; Williams, L. First Proof Second Batch. arXiv 2026, arXiv:2606.18119. [Google Scholar] [CrossRef] [Scilit]
- Zheng, D.; von Glehn, I.; Zwols, Y.; Beloshapka, I.; Buesing, L.; Roy, D.M.; Wattenberg, M.; Georgiev, B.; Schmidt, T.; Cowie, A.; et al. AI Co-Mathematician: Accelerating Mathematicians with Agentic AI. arXiv 2026, arXiv:2605.06651. [Google Scholar] [CrossRef] [Scilit]
- Decrypt. OpenAI Says a Secret AI Model Cracked Hundreds of Open Math Problems in One Prompt—Mathematicians Want Receipts. 7 October 2026. Available online: https://decrypt.co/380366/openai-secret-ai-model-cracked-hundreds-math-problems-one-prompt (accessed on 8 October 2026).
- Azzouni, J. The Derivation-Indicator View of Mathematical Practice. Philos. Math. 2004, 12, 81–105. [Google Scholar] [CrossRef] [Scilit]
- Tanswell, F.S. A Problem with the Dependence of Informal Proofs on Formal Proofs. Philos. Math. 2015, 23, 295–310. [Google Scholar] [CrossRef] [Scilit]
- Andersen, L.E. Acceptable Gaps in Mathematical Proofs. Synthese 2020, 197, 233–247. [Google Scholar] [CrossRef] [Scilit]
- Tao, T. Mathematics in the Age of AI. arXiv 2026, arXiv:2608.16753. [Google Scholar] [CrossRef] [Scilit]
- Avigad, J. Is Mathematics Obsolete? arXiv 2025, arXiv:2502.14874. [Google Scholar] [CrossRef] [Scilit]
- Avigad, J. Mathematicians in the Age of AI. arXiv 2026, arXiv:2603.03684. [Google Scholar] [CrossRef] [Scilit]
- Commelin, J.; Jamnik, M.; Ochigame, R.; Taelman, L.; Venkatesh, A. Shaping the Future of Mathematics in the Age of AI. arXiv 2026, arXiv:2603.24914. [Google Scholar] [CrossRef] [Scilit]
- International Mathematical Union. IMU News 137. 2026. Available online: https://www.mathunion.org/fileadmin/IMU/Organization/imu-net/pdfs/2026/IMU-News-137.pdf (accessed on 7 October 2026).
- Ioannidis, J.P.A. Why Most Published Research Findings Are False. PLoS Med. 2005, 2, e124, Correction in PLoS Med. 2022, 19, e1004085. [Google Scholar] [CrossRef] [Scilit] [PubMed]
- Woodward, J. Making Things Happen: A Theory of Causal Explanation; Oxford University Press: New York, NY, USA, 2003. [Google Scholar]
- Advisory Group on Mathematics and Artificial Intelligence. On OpenAI’s Release of Mathematical Results. 6 October 2026. Available online: https://agmai.org/ (accessed on 8 October 2026).
- Friedman, M.; Kish Bar-On, K. Rethinking Mathematical Intuition in the Age of AI. Math. Intell. 2026. [CrossRef] [Scilit]
- De Toffoli, S. Groundwork for a Fallibilist Account of Mathematics. Philos. Q. 2021, 71, 823–844. [Google Scholar] [CrossRef] [Scilit]
- Hardwig, J. Epistemic Dependence. J. Philos. 1985, 82, 335–349. [Google Scholar] [CrossRef] [Scilit]
- Avigad, J. Mathematics and the Formal Turn. Bull. Am. Math. Soc. 2024, 61, 225–240. [Google Scholar] [CrossRef] [Scilit]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license.