cert-machine · report · generated from the record

Twelve blocks on the integers: two AI models propose an exact minimum

I gave GPT-6.2 Astra (OpenAI) and Claude Fable 5.2 (Anthropic) one identical message and one shared folder, and asked what happens when the best known colouring for a question of Ron Graham’s, twelve blocks of red and blue, is put on the integers. In 73 minutes they proposed, and each read the other’s proof of, the exact least number of single-coloured 3-term progressions: at every length n = 548m with m even and at least 118, or odd and at least 123, and at every length from 92,000 on. With Claude (Anthropic’s AI model, and the same model as one of the two), I then decided the finite facts those proofs display: 10 statements, all certified, and 19 planted wrong versions, all refused. The theorem itself is not decided here. No machine has checked the proofs, nobody outside the two models has read them, and they rest on a lower bound that is this repository’s own and is not refereed.

0 4 8 12 0 55 110 165 220 274 n modulo 1096 (the first quarter of the period) progressions above 117n²/2192 − n/2 the twelve blocks rounded to the nearest integer: counted here from the colouring the minimum the two models propose for n ≥ 92,000: their table, not decided here
0 4 8 12 0 137 274 n mod 1096 above the formula twelve blocks, rounded (counted) proposed minimum (not decided)
What is left after the formula. The least number of single-coloured progressions is 117n²/2192 − n/2 plus a bounded correction. The solid curve is that correction for the twelve blocks rounded to the nearest integer, counted here from the colouring itself; it repeats every 1,096 lengths. The dashed curve is the minimum the two models propose for every n from 92,000 on: their table of 1,096 integers, which this page does not decide. Over the whole period it coincides with the rounded colouring on 500 residues and lies below it by at most 10 on the rest. The figure shows residues 0 to 274 of the 1,096; the record holds them all.Records: certs/twelve-blocks-series.json · sha256 b84437a3bddf; certs/twelve-blocks.json · sha256 46531914710fThe build refuses this figure if the proposed minimum rises above the counted colouring on any residue, or the counted series stops repeating with period 1,096.
finite statements decided
10 of 10
all certified, each on a stated finite domain; 19 planted wrong versions refused
proofs checked by a machine
0
each proof was read once, by the other model; one finite lemma is checked in Lean
least scale the proposed theorem covers
m = 118
n = 548m with m even; odd m from 123. Every other scale is open
from the first message to the last
73 min
47 messages between the two models, in one shared folder

This page reports one run of two models working together; it does not compare them. The mathematics is theirs. The checks and this page are mine, done with Claude, and the session that ran the checks is the same model as one of the two.

read this first

What is still open

The conditions of the run:

the question

How few single-coloured progressions can a colouring have

Colour each of the integers 1 to n red or blue and count the triples a, a + d, a + 2d, with d ≥ 1, whose three members have one colour. Call the least possible count V(n). A random colouring has about n²/16. In 1999 Ron Graham offered $100 for the true constant.

In 2008 Parrilo, Robertson and Saracino found a colouring in twelve alternating blocks, of lengths proportional to

28 · 6 · 28 · 37 · 59 · 116 · 116 · 59 · 37 · 28 · 6 · 28 (out of 548)

that has (117/2192)·n² of them to first order, and conjectured that nothing does better. In October 2026 this repository published a proof of the matching lower bound, by five exact certificates; it is its own report, and it has not been refereed.

That result is about the leading term. On the integers, with n = 548m and every block m times as long, the twelve blocks have exactly 117n²/2192 − n/2 single-coloured progressions when m is even, and one more when m is odd. The question I put to the two models: is that the minimum, exactly, once m is large?

the run

One message, one folder, two models

1 · what they proposed

Six claims, each read by the other model

PROOF READ means a written proof was followed line by line by the reader named and no error was found. It is not a formal check, and it is not this repository’s verdict.

ClaimFromRead byStandingHow it was read
In the 22-site family (two sites at each of the eleven block boundaries), the cost of every pattern of flips is the same at scale m and at scale m + 2, for every m.GPT-6.2 AstraClaude Fable 5.2PROOF READInherited from an earlier GPT session’s appendix. Claude derived the coefficients from the definition before opening the code, recounted all 253 with no formula at eight scales, and redid the affine step symbolically.
For every scale m, no pattern of flips inside the eleven windows of radius 2m around the boundaries (44m sites) goes below the twelve blocks. At odd m exactly 51 patterns tie, all inside the 22-site family; at even m only the empty one.GPT-6.2 AstraClaude Fable 5.2PROOF READClaude checked it from its own derivation, fixed before it read the proof: a positive definite quadratic polynomial in 22 integers whose non-positive points it enumerated.
For n = 548m with even m ≥ 118 or odd m ≥ 123, V(n) = 117n²/2192 − n/2 + (m mod 2), and the minimisers are exactly the twelve blocks and the tying patterns of the window theorem, with their colour exchanges: 2 at even m, 102 at odd m.GPT-6.2 Astra and Claude Fable 5.2Claude Fable 5.2, GPT-6.2 AstraPROOF READGPT proved it from m ≥ 192 (even) and 195 (odd). Claude’s sharper bound on what one flip can recover, with GPT’s deletion argument, gave 128 and 131; Claude’s constant 1/948 gave 118 and 123. Each read the other’s step.
At every length n the twelve blocks rounded to the nearest integer have 117n²/2192 − n/2 plus a term that depends only on n mod 1096 and lies between −369/548 and 1453/137.GPT-6.2 AstraClaude Fable 5.2PROOF READClaude recounted every length up to 3,288 by counting triples and read the proof.
At every length n, 117n²/2192 − n/2 − 281/16 ≤ V(n) ≤ 117n²/2192 − n/2 + 1453/137. At every scale m, the twelve blocks are within 14 (even m) or 16 (odd m) of the minimum.Claude Fable 5.2GPT-6.2 AstraPROOF READThe upper constant is GPT’s (ROUNDED); Claude’s own was 54. GPT rebuilt the certificates with its own program and re-derived the bounds.
For every n ≥ 92,000, V(n) = (117n² + c_r)/2192 − n/2, where c_r depends only on r = n mod 1096: 1,096 integers between −1476 and 3516, with the number of minimisers for each.GPT-6.2 AstraClaude Fable 5.2PROOF READClaude computed the table of 1,096 values with a program written before it opened GPT’s proof; the two tables agree on every row, in value and in number of minimisers.
2 · the argument

Four steps, and who took each

The same four steps, with the rounded colouring in place of the twelve blocks, give the bounds at every length and the table of 1,096 corrections. The proofs are corpus/twelve-blocks/run/gpt/global.md, corpus/twelve-blocks/run/gpt/windows.md, corpus/twelve-blocks/run/gpt/sharper-cutoffs.md, corpus/twelve-blocks/run/gpt/rounding.md, corpus/twelve-blocks/run/gpt/periodic-optimum.md and corpus/twelve-blocks/run/claude/NOTES-alln.md.

3 · what is decided here

Ten finite statements, and the wrong versions that had to fail

  • What this is. A judge that shares no code with either model decides the finite facts the proofs display, each on a stated finite domain, in exact arithmetic. The statements were written down before the judge existed (notes/twelve-blocks-preregistration-2026-11-19.md). It does not decide the theorem: a statement about every n is checked on the lengths named and no further.
  • The controls. 19 wrong versions of the same statements were planted and each had to be refused: the constant 947, the identity without its correction term, a bound one larger, cutoffs two smaller, a tie count of 50, a period of 548, and others. None was missed.
  • A fault of mine, kept. The first run refused one statement, D9. The fault was in how I had written it down: I had guessed the sign of one field of the models’ table. The statement was restated in a dated amendment before the judge was changed, and the first verdict is kept in the record.
What was decided, and on whatVerdict
D1The strengthened certificate, with the constant 948, is accepted by this repository’s own verifier, unmodified; 947 is refused.CERTIFIED
D2The identity on the integers, and the bound on its correction: every colouring of every length from 2 to 14, and 200 random ones at each length to 80 (45,964 colourings).CERTIFIED
D3A flip far from the boundaries costs at least 2m + 2, or 2m + 1 at odd m: every such site at m = 1 to 12 and at the two cutoffs, 118 and 123. The bound is attained.CERTIFIED
D4The arithmetic that turns the bounds into the cutoffs 118 and 123 and into the constants of the every-length bounds, by exhaustion over a box that contains every case.CERTIFIED
D5The twelve blocks have exactly 117n²/2192 − n/2, plus one at odd m, at 28 scales up to m = 195; among the 4,194,304 patterns of the 22 boundary sites none goes below them, with 51 ties at odd scales and none at even scales (m = 1 to 8).CERTIFIED
D6114 pieces of the wider windows, 24 sites each, walked in full at m = 1, 2 and 3: none goes below the twelve blocks, and the ties are exactly the expected ones.CERTIFIED
D7The rounded twelve blocks, counted at every length from 1 to 3,288: the excess repeats with period 1,096 and lies between −369/548 and 1453/137.CERTIFIED
D8The bounds at every length contain the 45 known values of V(n) (OEIS A121385); the first 14 were recomputed here by exhaustion.CERTIFIED
D9The table of 1,096 corrections, as a table, agrees on every row with the count of the rounded colouring and the loss the table states.CERTIFIED
D10At 24 lengths from 92,000 on, one in each of 24 residue classes, the colouring the table names has exactly the count the table states. This shows V(n) is at most that value there, and nothing about the other direction.CERTIFIED
The 19 planted wrong versions
PlantedWhat the judge foundRefused
K1 the constant 947947: REFUSED (psd: P is not proved PSD — exact witness vᵀAv < 0 at pivot 129 (rescaled value -4.7149e-3))yes
K2 the identity without Rfails on some colouring of length <= 8yes
K2 the identity with -Rfails on some colouring of length <= 8yes
K2 the identity d = 0 admittedfails on some colouring of length <= 8yes
K2 the identity T unroundedfails on some colouring of length <= 8yes
K3 the bound 2m+3-othe least cost equals 2m+2-o at m = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 118, 123]yes
K4 the maxima 236 and 243the maxima are 237 and 244yes
K4 the cutoffs 116 and 1212*116+2 = 234 < 237 and 2*121+1 = 243 < 244yes
K5 the odd tie count 50the census finds 51yes
K5 the zero mask unique at every scaleat m = 1 there are 51 minimisersyes
K5 the first-boundary swap costs m (the run’s plant)at m = 4 the swap costs +2yes
K5 the pair weight without the midpoint (-2 for sites 28, 34)the mixed term is -3; third sites [22, 31, 40]yes
K6 a piece’s tie count raised by onethe first piece has exactly 17 minimisersyes
K7 period 548fails first at n = 1yes
K7 the maximum 1452/137the maximum is 1453/137yes
K8 V(n) >= f(n) for every nV(n) < f(n) at n = [18, 38]yes
K9 a witness with its first listed site droppedresidue 548, n = 92612: M = 457757965 against 457757964yes
K9 a row with c_r lowered by 2192every lowered row disagrees with the count of C_nyes
KL the planted Lean file (period 1; one constant moved; the pair weight without its parity bracket)lean exits 1 with 4 errorsyes
What the judge does not decide, by name
  • the centring identity and E >= L (paper lemmas of the delta-three record)
  • that the identity verified in D1 yields E >= mu^2/948 (it does through those lemmas)
  • the lattice identity beyond the lengths of D2
  • the outside-window bound beyond the scales of D3
  • WINDOW beyond the pieces of D6: no full 2^(44m) family is walked, at any m
  • the localisation and last-deletion arguments of GLOBAL and SHARP
  • the affine and polynomial-in-k arguments of H3, ROUNDED and PERIOD
  • the lower bound of V in every statement: no V(n) is computed beyond n = 14
  • the multiplicities of PERIOD
  • that the two tracks’ programs do what their notes say
  • L1: that the cost of a pattern of flips is the quadratic form with the closed-form coefficients, for every m (the bridge from the definition of M stays on paper); no other lemma is in Lean
4 · one lemma in Lean

The step from two scales to every scale

prior work

What was already known

A web search of about thirty pages, not a survey. Nothing on this page should be called new before someone who knows this literature has read it. The constant 117/2192 itself is, in print, still a conjecture.

Not found: an exact value of V(n) beyond n = 45; an exact extremal statement for infinitely many n; the coefficient −1/2 of the second term, or any proved lower bound at that precision; a correction periodic in n. The search and its limits are in corpus/twelve-blocks/literature-2026-11-19.md.

who did what

Each step is a separate job

check it yourself

How to re-read it

contact

Write to Carlos

Sent by Web3Forms to Carlos’s inbox; nothing stored here.

Or write to carlos@carlostoledo.co, or open the public pilot form on GitHub.