Fifty new exact values for a colouring problem, each with a certificate
Colour the numbers 1 to n red or blue: how few three-term progressions of one colour can there be? The exact answer was known up to n = 45. In a second run, Claude Fable 5.2 (Anthropic) and GPT-6.2 Astra (OpenAI), working together, produced a certificate for every n up to 95: 50 new values. With Claude (Anthropic’s AI model, and the same model as one of the two), I checked every certificate in exact arithmetic with a program that shares no code with the search. These values depend on nothing but the definition. One further value is decided far beyond the table: at n = 548, the first length the twelve blocks of the 2008 conjecture fit exactly, the minimum is 15,756, attained by exactly 102 colourings. The two models also proposed much more, an exact formula for every large n; that part is not decided here.
What is left of the minimum after its two leading terms. Solid: the exact minimum V(n), decided here for every n up to 95 by a certificate and a colouring. Dashed: the formula the two models propose for every n, a table of 1,096 corrections that repeats with n, drawn where nothing is decided. On all 95 decided lengths the formula gives exactly the decided value. Beyond 95 it is a proposal: it is proved only from n = 15,756 on, by an argument this page does not decide.Records: certs/twelve-blocks-exact.json · sha256 87ded8bc3af5; corpus/twelve-blocks/run/gpt/optimum-residues.json · sha256 dbe25731effdThe build refuses this figure if a decided value leaves the proposed formula, or the table of decided values has a gap.
lengths with an exact minimum
95
50 of them new; the published table ended at 45, and its 45 values are reproduced
certificate pieces checked
33,368
each one a matrix proved positive semidefinite in exact arithmetic
planted wrong versions refused
14 of 14
a value raised by one, a piece dropped, a flipped site, and others
the exact minimum at n = 548
15,756
with 102 minimising colourings; no paper lemma used. The theorems about every large n are not decided
Two kinds of statement are on this page and they are kept apart. The exact values are decided, by certificates anyone can check. The theorems about every large n are the two models’ and are only read. Nothing here compares the two models.
read this first
What is still open
The theorems about every large n are not decided here. Each was read once, by the other model, and no reader outside the two model families has read any of it. They rest on δ₃ ≥ 117/2192, which is this repository’s own claim of October 2026 and is not refereed.
Between n = 96 and n = 15,755, at lengths that are not multiples of 548, nobody has a lower bound that matches the formula. A search from random colourings at 648 of those lengths found nothing below it; that is a search, not a proof.
The certificates slow down as n grows: about 3,000 cubes and twenty minutes of search per length near n = 95. The method does not reach n = 548 by itself: one cube proves 15,385 there and exactly no more.
The number of minimising colourings is decided only at n = 548. The table of exact values proves values, not counts.
The checker of the certificates was written by the control room, which is the same model as the track that found them. It is short, exact, and written from the format’s one page; the other model re-verified seven of the lengths with its own program.
The conditions of the run:
Both models worked in one shared folder and could read each other’s files and the whole first run from the first minute. This was a collaboration, and nothing here compares the two.
The control room that registered their claims, wrote the checkers and wrote this repository’s judge is Claude Fable 5.2, the same model as one of the two.
The checkers were written during the run and gave their verdicts there first. The pre-registration of the record was written after the run and says so.
The control room planted no false claim in this run. The only planted errors are the controls of the judge.
the certificate
How a lower bound is proved without trusting the search
One identity. Write +1 for red and −1 for blue. Three numbers have one colour exactly when (1 + ab + bc + ac)/4 is 1, and it is 0 otherwise. Summed over all progressions, four times the count is a fixed number plus a sum over pairs of sites, each pair weighted by how many progressions pass through both.
Cubes. The 2ⁿ colourings are split into disjoint pieces, each fixing the colour of some sites and leaving the rest free.
One matrix per piece. On each piece the certificate gives a list of integers. They define one integer matrix; if that matrix is positive semidefinite, the count on the whole piece is at least a number that can be read off, and one strict inequality makes it reach the claimed value.
The checker. It rebuilds each matrix from the piece and the definition, proves it positive semidefinite in exact integers and fractions, checks the inequality, and checks that the pieces are disjoint and fill everything. It was written from the format’s one page (corpus/twelve-blocks-2/run/claude/CERT-FORMAT-V.md) and shares no code with the program that searched. A certificate is checked, not trusted.
The other half. For each n a colouring with exactly that many single-coloured progressions is given and counted. Lower bound and colouring together are the exact value.
1 · the exact values
Every length from 1 to 95
What was known. The exact minimum for n ≤ 45 (OEIS A121385; the last values by integer programming, R. Pratt, 2014), and upper bounds 90, 95, 100, 104, 110 for n = 46 to 50.
What is decided here. The exact minimum for every n from 1 to 95. The five upper bounds of 2014 are sharp. The 45 published values are reproduced by certificates that use none of the earlier computations.
A pattern. Every one of the 95 values equals the formula the first run proposed for large n: (117n² + c)/2192 − n/2, with c one of 1,096 integers that depends only on the remainder of n modulo 1096. Whether that holds for every n is open; the two models conjecture it.
How far it goes. The largest certificate has 3,146 pieces. The model that found it reports about twenty minutes of search; checking all 95 certificates takes about two minutes.
All 95 values, with the number of pieces in each certificate
n
Exact minimum V(n)
Pieces in the certificate
1
0
2
as published
2
0
2
as published
3
0
2
as published
4
0
2
as published
5
0
2
as published
6
0
2
as published
7
0
2
as published
8
0
2
as published
9
1
2
as published
10
1
2
as published
11
2
2
as published
12
2
2
as published
13
3
2
as published
14
4
4
as published
15
5
2
as published
16
6
2
as published
17
7
2
as published
18
8
2
as published
19
10
2
as published
20
12
4
as published
21
14
12
as published
22
16
18
as published
23
18
24
as published
24
20
22
as published
25
22
16
as published
26
24
18
as published
27
26
8
as published
28
28
4
as published
29
31
8
as published
30
34
28
as published
31
37
42
as published
32
40
80
as published
33
43
70
as published
34
46
84
as published
35
49
54
as published
36
52
40
as published
37
55
14
as published
38
58
18
as published
39
62
36
as published
40
66
46
as published
41
70
80
as published
42
74
106
as published
43
78
114
as published
44
82
58
as published
45
86
64
as published
46
90
32
new
47
95
76
new
48
100
188
new
49
104
76
new
50
110
408
new
51
114
132
new
52
119
156
new
53
124
96
new
54
130
186
new
55
135
276
new
56
140
228
new
57
146
192
new
58
151
146
new
59
157
170
new
60
163
204
new
61
169
234
new
62
174
146
new
63
181
258
new
64
187
168
new
65
194
388
new
66
200
190
new
67
207
250
new
68
214
560
new
69
220
360
new
70
228
1,194
new
71
234
396
new
72
241
320
new
73
248
354
new
74
256
526
new
75
263
290
new
76
270
228
new
77
278
482
new
78
286
618
new
79
294
718
new
80
302
856
new
81
310
648
new
82
318
568
new
83
326
374
new
84
335
842
new
85
343
826
new
86
352
784
new
87
361
952
new
88
370
1,288
new
89
379
2,000
new
90
388
2,114
new
91
397
2,254
new
92
406
932
new
93
416
2,336
new
94
425
2,094
new
95
435
3,146
new
2 · the first twelve-block length
What is decided at n = 548
Why 548. It is the first length at which the twelve blocks of Parrilo, Robertson and Saracino fit the integers exactly: blocks of 28, 6, 28, 37, 59, 116, 116, 59, 37, 28, 6 and 28 sites. They have 15,756 single-coloured progressions. The first report is about whether that is the minimum at this and larger lengths.
A lower bound from nothing. One certificate with a single piece proves that every colouring of 1 to 548 has at least 15,385. A second exact computation shows that a single piece cannot prove 15,386: closing the remaining 371 this way needs many pieces.
Near the block boundaries. Among the colourings that agree with the twelve blocks outside 48 sites (four at each of the eleven boundaries, and four more), none has fewer than 15,756, and exactly 51 have 15,756. They are the same 51 found in the first report by trying all 4,194,304 patterns of 22 sites.
Everything else. A separate program, written from the text of GPT-6.2 Astra’s finite proof and sharing no code with either model, establishes in nine steps that a colouring with at most 15,756 must agree with the twelve blocks, or with their colour exchange, outside those 48 sites. Together: V(548) = 15,756 exactly, with 102 minimising colourings. Its report, gaps first, is instruments/twelve-blocks/exact/judge548/REPORT.md.
What that rests on. No theorem about the continuum and no lemma on paper: the inequalities of the October certificates are proved again by counting on the 548 sites. It does reuse the multipliers of those certificates as numbers, verifying every identity they enter, and it takes four of their bounds from this repository’s own verifier. One program, one author, the same model family as one of the two; nobody has replayed it with a second program.
Its planted errors. 8 wrong versions were planted; 7 were refused. One was not, and correctly: I had planned to double one entry of one of the proof’s 201 auxiliary vectors, but any such vector gives a true inequality, and that one still closes the argument. The control was badly chosen, and the record says so. Doubling the largest entry of each vector in turn stops the proof in 191 of 201.
3 · proposed, not decided
What the two models claim about every large n
PROOF READ means a written proof was followed by the reader named and no error was found. It is not a formal check, and it is not this repository’s verdict. All of this rests on the lower bound of October 2026, which is this repository’s own claim and is not refereed.
Claim
From
Read by
Standing
How it was read
For every n ≥ 20,700, a colouring with no more single-coloured progressions than the rounded twelve blocks differs from them only within 75 sites of the eleven block boundaries.
GPT-6.2 Astra
Claude Fable 5.2
PROOF READ
A new argument: restore one cell and bound the restored colouring. Claude rebuilt the key lemma along its own code path, re-derived the others on paper, and named the weakest steps: everything leans on one paper lemma of October, and three margins at n = 20,700 are within half a site.
At every length n = 548m, m ≥ 1, the twelve blocks are optimal: V(n) = 117n²/2192 − n/2 + (m mod 2), with 2 minimisers at even m and 102 at odd m.
GPT-6.2 Astra
Claude Fable 5.2
PROOF READ
The first run reached m ≥ 118. Claude replayed the finite parts with its own program. It did not re-check the first run’s classification inside the windows, which the result needs at every m ≥ 2.
V(548) = 15,756 with 102 minimisers, by finite checks on the 548-site model alone: no paper lemma and no continuum theorem; the October multipliers are reused as numbers whose identities are verified.
GPT-6.2 Astra
Claude Fable 5.2
PROOF READ
Claude rebuilt the two finite conditions that replace the paper lemmas with its own program; both hold exactly on both grids, one with a smallest slack of exactly zero.
Two explicit maps on the places where the colour changes: one enlarges a colouring by 1,096 sites and adds exactly 117n + 63,568 progressions, one sends length n to 1096q − n. Both keep the correction to the formula, which explains its period 1,096 and why it agrees at the residues r and −r.
GPT-6.2 Astra
Claude Fable 5.2
PROOF READ
Claude wrote its own test from the document and ran 1,880 cases with no failure, and showed the test fails outside the stated range.
For every n ≥ 15,756, V(n) = (117n² + c_r)/2192 − n/2 with the first run’s table of 1,096 corrections and numbers of minimisers.
GPT-6.2 Astra
Claude Fable 5.2
PROOF READ
The first run needed n ≥ 92,000. Claude redid the arithmetic in fractions and tested the maps at the smaller window.
A test that passes decides the test, not the theorem.
The formula’s value is attained. For every n from 1 to 3,288 some colouring has exactly the formula’s value. So the minimum is at most the formula there; the question is only whether it can be lower.
The two maps. At 40 lengths from 20,700 on, on three kinds of colouring each, moving every colour change by its prescribed amount adds exactly 117n + 63,568 progressions, and the reflection keeps the correction: 120 of 120 cases each. A shift one site larger gives the right number in 10 of 120.
The table’s colourings. At the first length from 20,700 on in each of the 1,096 residue classes, the colouring the table names has exactly the table’s value.
who did what
Each step is a separate job
The certificates and their format. Claude Fable 5.2, which also wrote the search that found them.
The theorems about large n, and the single-piece bound at 548. GPT-6.2 Astra. Claude Fable 5.2 read each of them; GPT-6.2 Astra re-verified the certificates at n = 46 to 50, 64 and 88 with its own program.
The checker and the judge. Written during the run by the control room from the format’s page, then pre-registered and run again for this record. It decides what it is given; it found nothing.
The run and this page. Carlos Toledo, with Claude (Anthropic), which wrote the programs under my direction. The run took 74 minutes on 19 November 2026 (UTC), on one laptop: 42 messages between the two models, in one shared folder.
check it yourself
How to re-read it
The record.python3 instruments/twelve-blocks/exact/judge.py --check re-checks the pinned files and decides four statements again in half a minute; --write decides everything, in about twelve minutes on eight cores, and writes certs/twelve-blocks-exact.json. The statements were written down before the judge was committed: notes/twelve-blocks-2-preregistration-2026-11-19.md.