cert-machine · the claim that ℤ is Diophantine in ℚ
Defining the integers in the rationals: the finite parts decided, the theorem not
On 5 November 2026 Elliot Glazer, working with AI agents, claimed one polynomial with integer coefficients, in 292,890,231,905 variables, whose rational zeros pick out exactly the integers. If it is right, a conjecture of Mazur is false and Hilbert’s tenth problem over ℚ has a negative answer. Everything in the claim a machine can decide today was decided again here, with exact arithmetic and code that shares nothing with his. The theorem itself is undecided.
since these checks
Versions 3 to 16: what moved, and what is not checked
New, and not checked here. Versions 14 and 15 were merged on 10 November. Version 14 gives quartics in 299 variables, with symbolic coefficients, and in 314 variables, with every coefficient in decimal (his counts). Version 15 selects a family of quartics in 222 variables; its README says the family’s defining threshold remains ineffective, that is, asserted to exist with no value given. Version 16 is an open pull request, not merged when this was checked on 11 November: by its description, three corrections to the manuscript after a further audit, with the constructions unchanged. None of this is checked here: not the quartics of Versions 14 and 15, not their manuscripts and review archives (8,865 files added to the repository since the Version 13 head), not the corrections of Version 16 (48 more files).
Before that. Between 6 and 8 November the repository moved from Version 3 to Version 13. Versions 3 to 7 (6 November) each brought a new manuscript and a far smaller explicit polynomial (a quartic) with its own package of checks. Versions 9, 12 and 13 brought smaller quartics again (by his counts 523, 498 and 394 variables, each with a companion that writes every coefficient in decimal: 540, 514 and 411); Versions 8, 10 and 11 changed the text and repaired proofs, Version 10 answering an outside review. The repository keeps the original large polynomial and the Version 4 quartic, decided in its own section below. The quartics of Versions 3, 5, 6, 9, 12 and 13 are not checked here either.
What still applies. Everything below was decided on the 5 November release. Three parts of the repository are the same bytes at the Version 15 head of its main branch and at the head of the Version 16 pull request: the original large construction and the whole Lean formalization, 86 of 88 files compared one by one, the two differences READMEs; and the Version 4 quartic’s package, 353 of 353 files. The Lean README says what the frozen target is: the statement for the original polynomial, not for any quartic, none of which has a Lean statement yet. So the Lean replay, Track B and the bookkeeping of the original construction apply to the original polynomial as the repository holds it at Version 15, and would still apply if Version 16 is merged as it stands (corpus/zq/version15-tree-check.json, corpus/zq/version16-tree-check.json).
the story in one minute
What was claimed, and what I checked
The claim. Hilbert’s tenth problem asks for an algorithm that decides whether a polynomial equation with integer coefficients has a solution. Over the integers there is none (Matiyasevich, 1970); over the rationals the question is open. Glazer’s manuscript and repository, github.com/elliotglazer/diophantine-z-q, claim the step that would close it: a fixed polynomial whose rational solutions, read in their first coordinate, are exactly the integers. Then no algorithm can decide rational solvability either.
What is in the repository. The polynomial is not written out as a list of terms. It is saved as a recipe of templates that call one another, from which its size can be counted. The proof is a manuscript of arithmetic geometry resting on published theorems. In Lean, a program that checks proofs line by line, 48 helper modules are accepted, and the main statement is there, frozen, with its proof left as sorry: unproved.
What I did. With Claude agents, I wrote down, before any of my code ran on his data, what each check would decide and what would make it fail. Then I decided again, in four lanes and with code of my own, every part of the claim that is finite: the number of variables and the height of the expression; the tower bounds on the degree and the coefficients; his four finite test suites; and his 48 Lean modules, replayed in Lean’s kernel on a rented machine.
What came out. The bookkeeping holds. The polynomial has exactly 292,890,231,905 variables, counted two ways, and its expression height is 256, as stated. The bounds on degree and coefficients follow by exact arithmetic, but only from four named premises; an attempt to remove one of them by computation was refused, because it is not a computation. The statements his suites test hold on 671,659 cases of mine, and his Lean helpers check in the kernel with only the three standard axioms. Then the Lean gap: a proof of his frozen main statement from eight named hypotheses, written with Claude agents in claimbox, checks in the kernel here, and one of his premises, that the point P₀ has infinite order, is now proved outright.
What it means. None of this decides the theorem. It decides what a machine can decide about the claim today and names what it cannot: the deep results the proof rests on, now eight named Lean hypotheses instead of a sorry. The limits come first below, then each lane with its record.
what stays open
What this does not settle
- The theorem is undecided here. The main Lean statement, lean/Frozen/BoundedIntegerPolynomial.lean, is unproved: its proof is sorry, and lane 4 found it elaborates with exactly one such sorry. Nothing on this page decides whether ℤ is Diophantine in ℚ, nor any consequence claimed for it (Mazur’s conjecture, Hilbert’s tenth problem over ℚ).
- The tower bounds rest on four premises. Lane 2 decides the exact arithmetic from them to deg P* < 2↑↑7 and |c| < 2↑↑8, and nothing more. H1, an explicit bound on primes in arithmetic progressions (Bennett, Martin, O’Bryant and Rechnitzer), as his proof applies it; H2, the inequality M0 ≤ Ns10, from his reduction-order and formal-group argument (Silverman); M, the degree and log-norm majorant at each expression height, a lemma of the manuscript, read and not decided; L1, the height 256, certified by lane 1. Without H1 and H2 the arithmetic gives no upper bound at all, so the bounds with no premise are REFUSED.
- H2 is not a computation. It is one inequality about fixed integers, so it is finite in principle, but deciding it directly needs every prime of his catalogue: 22·φ(W) of them, with φ(W) a 432-bit number. His proof of it instead applies three theorems at all of those places at once. Its nine finite ingredients are certified; the theorems are not, so H2 is REFUSED and lane 2 stays conditional on it.
- Lane 3 could replay his cases for one suite of four. His suites come with counts, but their cases are not in his prose: no ranges for the integer compiler’s digit, bit and packed cases (and its “backend” cases are not defined), no evaluation points for the division-polynomial graphs, which his code generates, and only sha256 hashes for the expression suite, with no stated serialization. Recovering them would mean reading his checker code, which a clean-room lane does not do, so those three replays are REFUSED. The statements each suite tests were decided on cases of mine instead; a count of mine that equals his is never presented as a reproduction of his set. The six-contact suite’s reported values were replayed, exactly.
- The Lean helpers check; what they mean is not checked here. The 48 modules check in the kernel, but whether their statements are the ones the manuscript needs is not decided. One second kernel, nanoda, could not finish the full export (a big-integer overflow, a limit of the tool), so the replay rests on leanchecker --fresh; nanoda accepted 47 of the 48 modules one by one.
- The Lean statement now rests on eight hypotheses. Track B decides that his frozen statement follows in Lean from eight named hypotheses (below). None of them is proved here. If any is false, the Lean theorem says nothing. Whether each Lean hypothesis says what the manuscript’s premise says is my reading, not a decision; a screen only checked that none of them contains his statement itself. Their joint consistency is not decided either.
- Some rules came after a first look. Every lane was pre-registered before any of my code ran on his data, and every later rule is dated in the pre-registration: lane 4’s rule for four constructors’ axiom sets was written after a first pass and before the recorded one, its per-module nanoda run during the recorded pass, and its recorder’s census fix after it. One red control of the H2 tool was blind for a structural reason, so a sharper one was added after the run. Lane 2’s red controls came after the pre-registration: a check of the tool, not a pre-registered test. In Track B, the rule for comparing the generator statement with mine (equal up to the names Lean gives its hidden helper proofs) was written after a first pass and before the recorded one.
- One author on both sides. The checks of Track A, the proof of Track B and its decision were all written by me with Claude agents. claimbox proposes and cert-machine decides; the two have one author, so this separates the roles, not the people. No one else has replayed these records yet; an outside replay would be the next step up.
the four lanes and track b
What was decided, lane by lane
| lane | verdict | what it decides | record |
|---|
| 1 · the count and the height | CERTIFIED | The ambient variable count, 292,890,231,905 = 1 + 292,890,231,904 witnesses, derived twice from the structure of his saved definition (the largest slot used, and every object’s own slots times its invocations), the slots tiling exactly at every object. The root expression height 256 ≤ 256 under the manuscript’s rules. His bounds file agrees on all 63 rows; nine red controls, each seen. | certs/zq-recount.json |
| 2 · the tower bounds | CERTIFIED conditional on H1, H2, M, L1 | deg P* < 2↑↑7 and every |coefficient| < 2↑↑8, by exact integer comparison over the constants of his saved definition with outward-rounded logarithms, from H1, H2, M and L1: 21 steps of the manuscript’s chain, each certified. Under the premises, log₂ log₂ log₂ of the degree bound is at most 8971; the claim needs less than 65,536. | certs/zq-tower.json |
| 2 · with no premise | REFUSED | The same two bounds, unconditionally: without H1 and H2 nothing bounds the catalogue or the formal multiple from above. | certs/zq-tower.json |
| H2 · removing a premise | REFUSED its ingredients CERTIFIED | M0 ≤ Ns10, refused as not a computation. Its nine finite ingredients, all certified: P0 on the curve, the places of bad reduction, the exact reduction orders of P0 at all 95 places above the 65 odd primes up to 324, the dyadic valuations and the exponent bookkeeping. Lane 2 stays conditional on H2. | certs/zq-h2.json |
| 3 · the finite suites | MIXED | His four suites re-derived clean-room from his prose and data, his checker code not read: the statements on 671,659 cases of mine, with 18 red controls; his own case sets replayed for one suite and refused for three (below). | certs/zq-finite.json |
| 4 · the Lean replay | CERTIFIED | The 48 accepted helper modules build and re-elaborate with kernel checking on. 990 declarations, his 704 and 286 private theorems, reach only propext, Quot.sound and Classical.choice, by two independent traversals. leanchecker, leanchecker-paranoid and leanchecker --fresh accept; nanoda accepts 47 of 48 modules. The frozen statement elaborates with exactly one sorry, and Mathlib’s hyperoperation 4 2 k = 2↑↑k is kernel-checked. Five planted wrong variants, each caught. | certs/zq-lean-replay.json |
| B · the Lean gap | CERTIFIED conditional on eight hypotheses | His frozen statement, his file unchanged, follows in Lean from the eight named hypotheses M, H, PC, C, LT, CC, H1, H2: the kernel accepts the proof three ways, and the statement is checked against his file twice. Details below. | certs/zq-trackb.json |
| B · P₀ has infinite order | CERTIFIED | The first half of his generator proposition, proved in Lean (claimbox’s 2-adic proof), with the point and the curve the manuscript’s own, number for number. | certs/zq-trackb.json |
| lane 3 · suite | its statements, on my cases | his own cases, replayed | my cases | red controls |
|---|
| the integer compiler | CERTIFIED | REFUSED | 670,930 | 8 |
| division polynomials over ℚ(√2) | CERTIFIED | REFUSED | 612 | 3 |
| finite polynomial expressions | CERTIFIED | REFUSED | 74 | 4 |
| the six-contact identities | CERTIFIED | CERTIFIED | 43 | 3 |
CERTIFIED: decided by exact arithmetic or by the kernel, as pre-registered. REFUSED: not decidable from these bytes, with the reason written in the record. MIXED: within one lane, some parts certified and some refused, each named. A conditional certification names its premises, and decides none of them.
track b
The Lean gap: his statement from eight named hypotheses
The question. Lane 4 shows that his Lean helpers check and that the main statement is still sorry. Track B asks a narrower question than the theorem: can that frozen statement, his file unchanged, be proved in Lean from named hypotheses, so that what stays open is a list of stated results instead of a sorry?
The proof. claimbox, the claimant of this machine’s Lean proposals, proved it: from eight hypotheses, each a deep result or a premise of the manuscript entering unproved, his frozen statement follows. Claude agents wrote it there; claimbox and this page have one author.
What was decided here. CERTIFIED, as pre-registered. Lean’s kernel accepts the proof three ways, the last replaying every constant it rests on, core and Mathlib included, into an empty environment (34 minutes). My audit finds only the three standard axioms in all 271 of claimbox’s declarations. The statement is his, checked twice: a Lean module of mine states it by taking the type of his frozen theorem from his file, and is proved by claimbox’s theorem; and a syntactic check finds exactly the eight hypotheses followed by his statement. Seven planted wrong variants, a changed conclusion, an extra hypothesis, a circular one, a sorry, a new axiom, a proof that skipped the kernel and a changed byte, were each caught. Every second, independent checker stopped on a resource limit before reaching a verdict (nanoda on a big-integer overflow, as in lane 4; con-ron and con-leche as Lean ships them on their computation budget; both of their master branches on a stack overflow, also on a 128 GB machine), so the kernel replay rests on Lean’s own checker, as lane 4’s does. On the planted variants nanoda did reach a verdict, and refused each for its planted reason.
One premise is now proved. That his point P₀ = ((1641 − 1160s)/8, (−123720 + 87483s)/32) has infinite order on y² = x³ − d²x with d = 75 − 53s, over ℚ(s) with s² = 2: the first half of his generator proposition. claimbox proves it at the prime 2, the way the manuscript argues elsewhere; the kernel accepts it, CERTIFIED, and the point and the curve in the Lean statement are the manuscript’s, number for number (CERTIFIED).
The eight hypotheses, in the order the theorem takes them, each in my reading, with where the manuscript has it:
- M. a polynomial of bounded size whose rational zeros are exactly the code triples of multiplication, (ι(a), ι(b), ι(ab)) (premise M, the multiplication section).
- H. the height bound h(u) ≤ H₆·M(u)²⁴ for every rational u outside the six contact points (the height theorem (thm:height)).
- PC. every rational B of the stated local shape is a value at which his pole polynomial has a rational zero (the pole hypothesis (hyp:PC)).
- C. a polynomial of bounded size that defines the precision machine’s closeness relation exactly (premise C, the compiler section).
- LT. a polynomial of bounded size carrying the six local tests, sound and complete (lem:test and lem:test-complete).
- CC. for every n outside the contact points and every κ ≠ 0, the precision machine accepts some run (eq:close-complete).
- H1. the explicit bound on primes in arithmetic progressions, as the cutoff proof applies it (proof of prop:cutoff (Bennett–Martin–O’Bryant–Rechnitzer)).
- H2. a formal multiple of P₀ of index at most Ns10 with the required divisibility, the same H2 as lane 2’s (proof of prop:cutoff (Silverman)).
What it does not mean. It does not prove that ℤ is Diophantine in ℚ: the theorem now rests on these eight, as lane 2’s bounds rest on theirs. If any of them is false, the Lean theorem says nothing. The readings above are mine, not decisions; the record keeps each hypothesis as Lean prints it. The plan this answers is this repository’s gap map, corpus/zq/gap-map.json, which named seven of them; building the proof found two more inputs (LT and CC) and proved one (G).
version 4
The Version 4 quartic, decided
What this does not decide. The theorem. The manuscript proves separately that its arithmetic source, a system of 4,513 equations in 4,935 rational coordinates, has a rational solution for exactly the integer inputs. That proof rests on the same inherited hypotheses as before and a sharper height estimate, and nothing here decides it. Nor whether those equations are the modules the manuscript describes, nor the short arguments in prose that each kind of step keeps the same inputs. It is not a Lean proof.
What it decides. The finite algebra that turns that source into one quartic, checked with my own code, written from the manuscript and the data files; none of his checkers was run or read. CERTIFIED: his coefficient table is a polynomial of degree 4 with 72,925 terms in 4,973 variables and largest coefficient 685,783,949,153,203,534,233,890, below 280, as stated. CERTIFIED: it is exactly the sum of the squares of his 4,552 published residuals, each of degree at most two, so at rational points it vanishes exactly when they all do. CERTIFIED: his certificate’s 551 steps (266 row operations, 148 eliminations modulo kept equations, 121 eliminations of a defined variable, 13 power circuits, 2 shears, 1 parity block), replayed from the source after the quadratic lift, each pass the condition the manuscript states for its kind and end at exactly those residuals.
So. If the source theorem holds, and the short step arguments are right, the quartic defines ℤ in ℚ. What stands between this quartic and the claim is now that one theorem and its inputs, not the algebra. All 13 controls came out as pre-registered: a changed coefficient, a dropped residual, a doubled weight, one planted error for each kind of step and the input deleted were each refused, and my arithmetic caught its planted non-identities. The pre-registration was written before any code ran; its four dated amendments each correct a reading of mine (a header field, the sign of a row), none a result (notes/zq-v4-preregistration-2026-11-06.md, certs/zq-v4.json).
check it yourself
How to check this yourself
- Get his repository at the pin. git clone https://github.com/elliotglazer/diophantine-z-q, then git -C diophantine-z-q checkout 97cec7fdd757d9a8e4f468151a47a6d3610fb2f2. Point ZQ_REPO at the clone (the default is ~/Projects/zq-scout/diophantine-z-q).
- Re-derive the records. In a clone of cert-machine, run node tools/zq-recount.js --check, node tools/zq-tower.js --check, node tools/zq-h2.js --check, node tools/zq-finite.js --check, node tools/record-zq-lean-replay.js --check. Each checks the sha256 of every file of his it reads against the pin before reading it, derives its record again and refuses on any difference. They need Node and its standard library, and his clone at the pinned commit.
- The Lean replay. It ran on a rented machine with tools/runpod-zq-lane4.sh; the facts of the run are in corpus/zq/runs/lane4-2026-11-06/. The recorder’s --check derives again from your clone what can be derived on a laptop (the census, the re-hash, the reconciliation with his 704 declarations) and refuses if a script that ran on that machine has changed since.
- The Lean gap. claimbox’s proof is public on its branch zinq-2026-11-06. It was decided on a rented machine with tools/runpod-zq-trackb.sh, the facts in corpus/zq/runs/trackb-2026-11-06/; node tools/record-zq-trackb.js --check derives the record again from those facts and the manuscript’s numbers, and refuses if a script that ran there has changed since. Its pre-registration: notes/zq-trackb-preregistration-2026-11-06.md.
- The Version 4 quartic. A sparse clone of his repository at 5cb2c199 (the package quartics/quartic-4973/ and quartic_4973.tex), then ZQ_V4_REPO=<clone> node tools/zq-v4.js --check, about three seconds on a laptop.
- Nothing of his is copied here. His repository has no licence (no licence file at Version 15; the README of the 5 November release said licences for original project material were awaiting the maintainer’s selection), so no file of it is in this repository or on this site. Every tool reads his files from your clone; the page quotes his statement only.
- Read the pre-registration. notes/zq-preregistration-2026-11-06.md: what each lane would decide and what would make it fail, written before any code ran on his data, with its dated amendments.