Notes from genuine attempts on open problems. Each one states the problem in plain language first, then goes as far as the mathematics actually allows — and says plainly where it stops.
Nothing here claims a proof of the headline conjecture. What a note does claim is checkable: exact values carry machine-verified certificates, bounds carry proofs, and anything fitted to data is labelled as a fit rather than a fact.
Above: the halved prime gaps under repeated absolute differences, parity only. Since |a − b| mod 2 is XOR, what appears is Pascal's triangle mod 2 — the Sierpiński gasket, seeded by the primes.
Put whole numbers on the edges of a tree and add them along every path between two vertices. With eleven vertices there are fifty-five paths, so fifty-five totals. How many of 1, 2, 3, … can you force to appear? John Leech worked this out to ten vertices in 1975, and the sequence database has carried "at least 48" for eleven ever since, with nobody named for it. The answer is 49: one tree reaches it, and an exhaustive search over 1.40 billion nodes shows that none reaches 50. Twelve vertices are now bracketed 57 to 60, where the database recorded 55 from below. A second sequence appears when the totals are required to be distinct, and that one is not in the database at all. The refutations at eleven and twelve vertices come from a single search program, and the page says so where the numbers appear.
Certified: a(11) = 49 First new term of A007187 since 1975 57 ≤ a(12) ≤ 60 1,400,728,816 nodes to refute 50 Single-engine at n = 11, flagged Eight colours for the triangular lattice, and a table that does not hold September 16, 2026Colour the points of a lattice so that no walk ever reads the same colour pattern twice in a row. How many colours does that take? Lower bounds come from finding one finite patch that cannot be coloured, and a paper from October 2025 pushed three of them up by counting colourings of growing patches until the count hit zero. Rebuilding that computation as a satisfiability problem with machine-checkable refutations gives a mixed verdict: the square-grid theorem is confirmed, and this page marks its own version as a rediscovery; the king's-graph bound is true but the patch offered for it is colourable, and the real threshold is three vertices later; and the triangular-lattice theorem has no valid proof, because the patch its table declares impossible can be eight-coloured, as the figure shows. New and certified: the triangular lattice needs at least eight colours, where the best valid bound was six.
Certified: π(T₃) ≥ 8 Published π(T₃) ≥ 9 has no valid proof Not refuted: the bound may still be true 194,113,939 paths checked on the disputed patch Square-grid bound marked as a rediscovery Seven letters, by way of the dragon curve September 9, 2026Walk on a grid, one unit step at a time along a positive axis, and never let three of the points you visit fall on a line. How many directions do you need? Erdős asked; the question was settled this month at sixteen, and the number has been coming down fast. Here the walk is the dragon curve, the shape a strip of paper makes when folded in half repeatedly and opened at right angles. Labelling each step by its direction and the turn that follows gives eight labels, two of which can be merged, and the resulting word has exactly the property the problem needs. Seven directions suffice, proved in a page. Seven is not the record: an unpublished two-page note reaches six, and this session read it line by line looking for a mistake and found none. What the dragon adds is an independent mechanism.
Proved: kmin ≤ 7 Not the record: a draft reaches 6 That draft reviewed, no gap found Word checked 3-free to 100,000 4 ≤ kmin still open Every seven positive reals contain a dissociated quadruple September 8, 2026Write down some numbers and add up every group of them. If no two groups ever share a total, the numbers are dissociated. Erdős asked in 1965 how large a dissociated group must hide inside any n numbers, guessed about log₂n of them, and sixty years later no exact value had been published. A one-line reduction collapses the question to sets of distinct positive numbers, halving the size and turning it into one finite statement for each size. The first open case then falls to an exact case analysis, machine-checked in rational arithmetic: seven numbers always contain a dissociated four, six need not. That fixes Erdős's function exactly up to 27, verifies his guessed inequality up to 31, and finds it strict at 14 and 15. The fifth threshold is not decided here, and the page says where the search stalled.
Certified: m₄ = 7 First exact values of f(n), n ≤ 27 Bound verified to n = 31, strict at 14, 15 Three case trees re-checked: 290, 111, 62 nodes 14 ≤ m₅ ≤ 41 undecided Sixteen vertices are enough to colour unevenly September 8, 2026Colour the dots and the lines of a network where every dot has three lines, so that things which touch never share a colour. Four colours often suffice; using them evenly, each about as often as the others, is a different matter. A paper posted on 4 September 2026 exhibited a 20-dot network that can be four-coloured but never evenly, and asked whether anything smaller does the same, noting that a computer could settle it over the catalogue of smaller networks. It can, and the answer is 16, not 20: among all 45,982 connected cubic graphs on at most 18 vertices, exactly 23 are exceptions, sixteen of them on 16 vertices. Every refusal carries a machine-checked proof, and the figure re-runs the search in your browser. Priority against an earlier dissertation is unchecked, and the page says so.
Certified: minimum order is 16, not 20 Exactly 23 exceptions below order 20 45,982 graphs swept, 4,920 DRUP proofs checked Every exception splits 11-10-10-9 Priority unchecked, flagged A third peaceable queens number: a(18) = 47 September 7, 2026Two equal-sized chess armies, one white and one black, share a board under one rule: no queen may attack an enemy queen. OEIS's recorded brackets for the largest army that fits had stood since 2014 at [37, 64], [42, 72] and [47, 81] on the 16, 17 and 18-square boards. All three are now closed: 37, 42, 47, each equal to the count of a construction published in 1977 that had never been proved best possible. The newest rung ruled out 48 per side by exhausting 119,110,352,726 search nodes, about nine core-hours, and the 47-per-side placement is checked here from the definition. The 16-square refutation was run by two independently written engines; the other two were run by one engine each, and the page states that as a defect rather than burying it.
Certified: a(16) = 37, a(17) = 42, a(18) = 47 Three smallest open cases of A250000, decided 119,110,352,726 nodes, 16/16 chunks UNSAT n = 17 and n = 18 single-engine, flagged All three equal ⌊7n²/48⌋ Chromatic Ramsey numbers: the growth rate of F(3,k) is exactly 3 September 6, 2026Colour every line between n people with k pens so that no three of them form a single-coloured triangle, and so that each pen on its own is simple enough to be sorted into three groups with no line inside a group. How large can n be? The obvious ceiling is 3k, and the best construction on record only reached about 2.83k. A half-page construction closes that gap: blocks indexed by subsets of the pen set, a binary cube inside each, and one sign compared across. The same argument gives rate exactly j for every fixed number of groups. Along the way, the first exact values anyone has computed: 14, 41, and 16, with machine-checked refutations of the next value up. Sawin's actual question, about the constant rather than the rate, stays open, and the page draws the gap.
Proved: lim F(3,k)1/k = 3 Certified: F(3,3) = 14, F(3,4) = 41 Previous rate 23/2 First exact values beyond F(2,k) Sawin's constant still open Fifteen points and seven ordinary lines: most of the case tree, closed September 5, 2026Scatter dots on a page and draw every line through at least two of them. A line through exactly two is called ordinary, and the Dirac-Motzkin conjecture says fifteen dots can never manage fewer than eight. Nobody knows whether that is true: fifteen is the smallest case still open, and the answer is somewhere in 7, 8, 9. This does not close it, and says so on every screen. What it does is cut the tree down: if fifteen dots really managed seven, their two five-point lines would have to meet, and between one and five of the sixteen pairs joining them would have to be ordinary. That took 895 solver refutations, every proof independently checked. Forty-five classes and about 25 processor-hours are left, and the page measures them rather than guessing.
Certified: disjoint case closed, 261 proofs Certified: meeting case with s = 0, 411 proofs 895 refutations, all drat-trim verified 45 classes open, ~25 CPU-hours t₂(15) still unknown Triangulation discrepancy: the open residue class, at its first four orders September 4, 2026Cut a shape into triangles and paint every corner red or blue so that no triangle comes out all one colour. How evenly can the two colours be split? A paper from August proves a sharp bound for every number of corners except one family, 11, 17, 23, 29 and so on six apart, which it leaves open after checking the smallest case by hand. This settles the next three: a structure theorem pins down exactly what a counterexample would have to look like, and that shape cannot exist below 29 corners. At 29 it is not proved but ground out, all 977 million triangulations on 18 corners reduced to 60 million candidate shapes, every one of discrepancy 1 against a bound of 7. The exact discrepancy table also goes from 12 corners to 17. The family itself stays open from 35 up, and the page says precisely where the argument gives out.
Proved: refined bound at n = 11, 17, 23 Certified: n = 29, 60,070,002 candidates Census to n = 17, all 129,664,753 Structure theorem for any counterexample Still open for n ≥ 35 The antidiagonal anomaly stops at 495, and never comes back September 4, 2026Walk across a square grid from one corner to the other, always east or north. Block one crossing on the anti-diagonal, and ask which nearby crossing carries the most surviving routes. Usually the answer is the one a step in from the edge, but for every grid up to 375 wide the winner flips to a point sitting flat on the boundary, then flickers on and off until 495 and stops. A paper five days old had verified this to 495 and conjectured it never happens again. It never does: a Stirling bound settles every size from 3000 up, exact integer arithmetic settles the 2504 in between, and the ratio that decides the question is shown to converge to √(8/π)e−1/2 = 0.967882. The flickering turns out not to be irregular at all, just the best blocking position failing to land on a whole number, with the even and odd sizes taking turns.
Proved: Conjecture 7.4 of arXiv:2609.01562 Certified 496 ≤ n ≤ 2999, exact integers Limit c0 = 0.967882… 455 anomalous sizes listed exactly Question (3) of §7 not addressed The Bit Deletion game, solved September 3, 2026Write a number in binary and take turns rubbing out one digit; whoever removes the last remaining one wins. Somebody computed the game's values to five million, noticed they never exceed 3 and never change when you multiply the starting number by four, and filed both observations in the sequence catalogue as conjectures. Both are true, and for a reason you can carry out by hand: count the digits for one bit of the answer, then count the odd-length runs of noughts before the first even-length one for the other. The closed form settles both conjectures, gives the exact count of positions of each value, and solves the reverse game where taking the last one loses. Who wins, as opposed to by how much, was already Project Euler folklore, and the page says so.
Proved: G(n) = (L mod 2) + 2·[t odd] Both A398916 conjectures settled Misère form solved Certified to n < 232, 0 mismatches Outcome rule marked as rediscovery 18 pseudolines, 93 triangles: a rediscovery, and an audit of the upper bounds September 2, 2026Draw eighteen lines across a page and count the triangles they cut out. Whether 94 is reachable had been marked undecided in the tables since 2007. A counting argument pins any 94-triangle arrangement to a very rigid shape, which turns an intractable search into 561 quick ones; all 561 are impossible, every one with a machine-checked proof. Then, mid-session, the author found that Blanc proved the same thing in 2008, so this is an independent confirmation and is labelled as one. What is new is uncomfortable: the upper bounds catalogued for this problem at even numbers of lines are theorems about lines in general position, while the record arrangements at 8, 12 and 14 lines provably are not in general position. On the cited literature the honest brackets are wider than the ones on record.
Certified: as3(18) = 93 Rediscovery of Blanc 2008, marked 561 cubes UNSAT, 561 drat-trim-verified 5.0 + 8.2 core-hours, 27.75 GB of proofs Audit flagged: 54 ≤ K(14) ≤ 55, 93 ≤ K(18) ≤ 95 The first upper bound for Λ(8,2), sixty-two years on September 2, 2026Fix a prime and hunt for the first two consecutive numbers that are both perfect eighth powers modulo it. How far can a prime force you to go? Brillhart and the Lehmers answered the same question for squares up to seventh powers in 1964, then hit eighth powers, where a forced fact about the number 2 broke their theorem. They left a lower bound of 1,200,744 and nothing above it. This note supplies the missing upper bound, 1,508,324, by finishing their case analysis on hardware they would have found absurd: three and a half million cases, each with a named pair of consecutive numbers that settles it, verified by an independently written checker. The exact value is still open, and every attempt here to raise the lower bound failed.
Certified: Λ(8,2) ≤ 1,508,324 Erdős #436, exact value still open First explicit bound at k = 8 3,499,913 leaves, none unsettled Λ(2…7,2) reproduced exactly A five-colouring of PG(7,2), if it exists, has almost no symmetry September 1, 2026Take the 255 non-zero strings of eight bits, call three of them a line when they XOR to zero, and try to paint them with five colours so no line comes out all one colour. Whether it can be done was asked in December 2025 and is still open. This note does not answer it; it removes the standard way of trying. Symmetric colourings, where whole orbits of a shuffle share a colour, are the usual route, and here every symmetry of odd order is dead: four orders fall to a one-paragraph argument that works in every dimension, and the rest fall to certified search, one of them needing a proof file of 812 megabytes. So any witness must be lopsided. One dimension down the symmetric colouring does exist, and the note ships it.
χ2(8) still open Proved: no Mersenne-order symmetry, any n and k Theorem: every witness has a 2-group stabiliser Certified: orders 17, 5 and Frobenius all UNSAT 166,270-lemma DRUP re-verified three ways 999 of 1,000 sampled witnesses distinct, 0 extend Graham's rearrangement conjecture, certified for every prime up to 37 August 30, 2026Write the numbers 1 to p−1 around a clock with p hours, choose some of them, and walk around the clock stepping by each in turn. Graham asked in 1971 whether you can always order your choice so that you never land on the same position twice. Since 2024 it is known for every large enough prime, but none of those proofs says how large, so the small primes are exactly what theory cannot reach. The published record was an uncertified 2016 random search that stopped at clock size 25. This session settles every prime up to 37 exhaustively, with an explicit good order for all 1,954,471,973 essentially different cases, and clears one awkward family out to 61.
Conjecture still open Certified: every prime p ≤ 37, all sizes 1,954,471,973 orbits, 70,066,181,009 subsets 0 failures, Burnside-exact on 173 cells Zero-sum p−3 layer certified to p = 61 Smallest undecided prime: 41 Counting the lattice points of the parking-function polytope August 29, 2026A one-way street with n spots, n cars, each driving to its favourite spot and taking the next free one if it is occupied. The lists of favourites where everyone parks are the parking functions. Treat them as points in space, stretch a skin around them, and count the whole-number points inside: at three spots there are seventeen, one more than the sixteen parking functions. Richard Stanley put the sequence in the encyclopedia in 2020 with eight terms, and two guesses about the answer were later attached to it. This note proves both, answers a separate question from a 2024 sandpile paper, writes down the polytope's full counting polynomial, and computes forty terms where eight were known.
Proved: both A333331 conjectures Proved: Ehrhart polynomial Answers Selig, EJC 31(3) 2024, section 6 Certified: a(1)–a(40), 8 were known a(9) = 167,341,283, never computed before Ratio to the proved law: 0.984 at n = 40 The smallest counterexample to the NCI conjecture has at least 16 elements August 28, 2026A solitaire game on a diagram: build a target set out of given sets, adding only sets that do not overlap what you hold and subtracting only sets that sit entirely inside it. Ordinary inclusion and exclusion is forbidden, because every intermediate position has to be an honest set. Three computer scientists conjectured in 2024 that the game can always be won; in August 2026 it was refuted twice, but both refutations only prove a losing diagram exists somewhere near 1015 elements and ask how small one can be. Nothing was known from below. This session checked every diagram with at most 15 elements, all 171,432,955 of them, and every single one is winnable.
Conjecture already refuted Certified: all 171,432,955 lattices to n = 15 winnable First lower bound on either open problem 34,978,238,589 posets streamed Counts match OEIS A000112 and A006966 at every size Gap remaining: [16, 1.00011×1015] A certified balanced five-colouring of 25 points, and no symmetric counterexample at 26 August 27, 2026Give every pair among 25 people one of five colours, so that any six of them show all five colours between them. It can be done, using the geometry of a 5 by 5 grid, and this note verifies it over all 177,100 ways of choosing six. Erdős and Gyárfás conjectured in 1999 that 26 is impossible for five colours, and nobody has settled it; this session maps the open case rather than deciding it. The obvious counting argument would close the problem if a certain extremal number stayed below a threshold, but it misses by exactly one edge in the already-proved three-colour case and by at least five at five colours, so it never can. The grid construction provably cannot be stretched to 26 points, and all 3,198 symmetric candidates fail, so any counterexample would have to be a lopsided object with no symmetry to help find it. The 26-point question itself stays open.
K26 still open Certified: balanced 5-colouring of K25 Proved: classes are (6,6)-Ramsey E*(10,4) = 31 against threshold 30 E*(26,6) ≥ 265 against threshold 260 0 of 3,198 symmetric candidates survive 1,160-line DRUP proof checked A diamond-free claw-free cubic graph that needs seven colours August 26, 2026Take a shape with three edges meeting at every corner, slice off each corner so it becomes a triangle, then colour the edges so that no two within one step of each other share a colour. Six colours are obviously the minimum, and in November 2025 František Kardoš asked at a workshop whether six always suffice. They do not: there is an 18-corner shape that needs seven, it is the smallest and the only one of its size, and a ten-line argument about a configuration called a balloon produces more at every size above it. But every counterexample turns out to have a weak point, a single edge holding it together, so the two readings of the question that most people would have meant both survive: all 26,867 cases with no such weak point and all 556,471 truncations of ordinary graphs tested are colourable with six.
Answered: no, as posed Certified: 18-vertex counterexample, χ′s = 7 Proved: the balloon obstruction Every counterexample is bridged 36,093 quotients decided, 7 ⇔ balloon exactly 556,471 simple truncations verified Odd Giuga and primary pseudoperfect numbers need at least 14 prime factors August 23, 2026Pick a few different prime numbers, add up one divided by each, then add or subtract one divided by their product, and try to land exactly on 1. Every solution anyone has ever found uses the prime 2; whether one exists built only from odd primes is open, and a counterexample to Giuga’s 1950 primality conjecture would have to be exactly that. This note exhausts both defining equations through 12 prime factors and rules out 13 by a parity argument, so an odd solution of either kind needs at least 14 different primes. For Giuga numbers that independently certifies a bound recorded in 1996 whose code was never published; for primary pseudoperfect numbers it moves the recorded bound from 9 to 14. Both published censuses, from 1996 and 2000, are reproduced from scratch term for term, and the wall that stops the search at 13 is measured rather than asserted.
Existence still open Certified: nothing odd below 14 factors Odd pseudoperfect bound 9 → 14 1996 & 2000 censuses reproduced 28,131,218,255 closures at m = 12 A sharp extremal length for undirected repetition thresholds August 18, 2026Write the longest string you can over a fixed alphabet so that no stretch of it repeats a pattern too aggressively, forwards or as a mirror image. For every alphabet of 5 or more symbols, this note proves the longest such string has length exactly three more than the alphabet size, with an essentially unique example achieving it: a new elementary proof feeding into half of a 2021 conjecture, still open at the specific alphabet size (22 letters) the session set out to decide. Also ships exhaustive certificates for four small cases, a 20,000-symbol witness that the space of legal strings stays enormous just past the threshold, and two wrong conjectures the session caught and corrected in itself before publishing.
k = 22 still open Proved: length n+3, unique word Certified k = 22–25 20,000-symbol witness Seven certified van der Waerden numbers, and a new bound past w(2;5,7) August 16, 2026Color the whole numbers from 1 to n in two colors and ask when you're forced into an evenly-spaced run in one of them. Seven cells of this table, computed between 1978 and 2013 with no certificate beyond "the program said so," get their first machine-checkable proofs here. The open cell just past the table's 2013 frontier gets its first recorded bound at all: w(2;5,8) > 295, from a coloring that repeats every 74 numbers, verified by two independent from-definition checkers.
w(2;5,8): still open above 295 w(2;5,8) > 295, certified 7 cells: first machine-checked proofs DRUP-verified, up to 18.4M lines Grimm's conjecture, certified for all n ≤ 1012 August 15, 2026Between consecutive prime numbers sit runs of composite numbers, and Grimm conjectured in 1969 that every run can be tagged with distinct prime factors, one per composite, no two composites sharing a tag. This note extends the computational record 52 times further, to 1012, and censuses every number that ever puts the tagging at risk, for the first time: 18.6 million of them, every one successfully tagged, and the 133 tightest cases all traced to the same simple reason: a single prime power sitting in the gap.
Still open Certified n ≤ 1012 52× the 2006 record 18,575,022-member census 133/133 tight gaps explained The census of Graham's $1000 problem, 216 orders of magnitude further August 14, 2026Ronald Graham offered $1000 for a proof of whether infinitely many n make “2n choose n” simultaneously indivisible by 3, 5 and 7. No computation can ever answer that, but this note pushes the complete list of known exceptions 216 orders of magnitude past the previous record, from 1,374 known cases to 585,823,270, and finds that the list arrives in bursts: a 27-decimal-order stretch with nothing in it, then thirty-nine million at once.
Still open (unresolvable by computation) Certified census to 3600 585,823,270 terms, was 1,374 216 orders of magnitude further A new certified floor for Erdős's distinct subset sums problem August 13, 2026Pick ten different whole numbers and add up every possible subset, 1,024 totals in all: a set has distinct subset sums if none of those totals ever tie. Erdős asked in 1931 how small the largest number in such a set can be forced to be — exact answers were known up to nine numbers, but not ten. Four cross-verified search engines push the certified floor on the tenth value from 220 to 262, narrowing the answer to the window (262, 309]; the known ladder for one through nine numbers is re-derived from scratch with every optimal set enumerated, and 19.1 million candidates near the best known construction are ruled out as a smaller witness.
f(10) still open f(10) > 262, certified Window narrowed to (262, 309] Full f(1)…f(9) ladder re-derived 19.1M candidates ruled out Signed difference sets: closing two-thirds of an open database August 9 & 12, 2026Mark each spot on a ring with a plus, a minus, or nothing, so that spinning a copy underneath and multiplying every lined-up pair of marks always sums to the same total, no matter the spin: a signed difference set. D. M. Gordon's 2023 database left 67,823 of these existence questions marked simply Open. Two classical nonexistence theorems, applied by computer to all 70,543 cells, plus a validated exhaustive search, close 45,370 of them in one session and turn up ten signed difference sets with no published source; an independent audit finds that 147 of the database's own 280 worked examples fail its own defining equation. A second session independently verifies an external researcher's closure of the last small open family (order 32 and 36) leg by leg, and adds a database-free proof for order 36.
22,453 cells still open 45,328 closed by two proofs 10 new signed difference sets 58 open cells decided by exhaustive search 147 of 280 stored witnesses invalid Order 32/36 family closed, verified Reciprocal Rado numbers: an independent certificate for nine exact values August 8, 2026Paint the whole numbers red or blue and ask whether you can always avoid a same-colored solution of 1/x₁ + … + 1/xₖ = 1/xₖ₊₁: past one exact threshold, every coloring fails. Two research groups had already computed nine of these thresholds this year, most recently one month earlier, with no proof beyond "the program said so." An independent, from-scratch pipeline with a DRUP proof on one side and a doubly-verified witness on the other reproduces all nine exactly, adds the extremal coloring structure neither prior source examined, and gives a candidate mechanism for half of a conjecture the source paper could only state from its own data.
Still open 9 values independently certified Matches Myers–Parrish & Gaiser–Ramezanpour Even-k mechanism for their Conjecture 1.3 f4(2) > 60000 Eleven new generalized Schur numbers, the first since 2016 August 7, 2026Paint the numbers 1 to m with three colors so that no color contains a monochromatic solution of its own forbidden-sum equation: past one exact m, every coloring fails. The published table of these thresholds — 26 exact values — had not grown since 2016. A SAT solver with a two-sided certificate (a checked proof on one side, an independently verified coloring on the other) finds eleven more: four confirm open cases of a 2016 conjecture, and seven map a previously unmapped family so cleanly that a new pattern predicted two of the new values before either was computed.
Still open 11 new certified values 4 confirm Ahmed–Schaal Conj. 2.1 Conjecture A predicted u = 10, 11 DRUP + independent witness, every value Circular repetition thresholds: two open cases settled Aug 3 & 6, 2026Bend a word into a necklace and ask whether you can always avoid over-repeating letters, at the same threshold that works for an ordinary word: open for necklaces of 4 to 44 letters. A first session built a pumping lemma and proved the natural growing rule impossible outright for 4–9 letters. A second session found a different growing rule one level of abstraction up — hidden in a binary code rather than the necklace's own letters — and it works: the 6-letter case is settled outright, the 5-letter case too (flagged as a likely overlap with a hard-to-read December 2025 paper), and the 4-letter case is now certifiably empty in this encoding as well.
Still open (n = 4, 7–44) CRTW(6) = RT(6), settled CRTW(5) = RT(5), flagged Pansiot-code pumping lemma n = 4 empty, two independent searches Consecutive powerful numbers in arithmetic progression: a census to 1019 August 5, 2026A powerful number never has a prime factor that shows up just once. Van Doorn conjectured in May 2026 that infinitely many triples of consecutive powerful numbers are evenly spaced, with just 18 such triples known below 1014. This note extends the census five orders of magnitude to 1019, finding 346 triples that collapse under an exact rescaling rule to only 16 genuinely different ones, including the first known triple with no perfect square in it, and proves that the easy route to infinitely many (rescaling one triple forever) provably cannot work.
Still open Census to 1019: 346 triples 16 primitives, 346/346 verified Proved: Lemma 1′ + Lemma 2 First squareless triple Finch’s regularity conjecture: certificates for 32 Ulam-type sequences August 2, 2026Build a list where each new number is the smallest one that is a sum of two earlier numbers in exactly one way: some starting pairs make the gaps between numbers repeat forever (regular), and Finch guessed which ones in 1991. A self-correcting finite certificate proves regularity for 32 starting pairs, 20 of them in cases open since 1995, filling exactly the gap the one existing 1995 proof left uncovered. Along the way: a hidden fourth even number that appears only when the second starting number is one less than a power of two, at a formula confirmed across 255 cases.
Still open 32 sequences certified regular 20 in cases open since 1995 Exceptional family at b = 2k−1 Self-correcting certificate Additive squares (PVHH): the singular relation August 1, 2026Split a list of numbers into two neighbouring equal-length pieces; if both add up to the same total, that’s an additive square. Fix a set of numbers and ask for an infinite list with no additive square anywhere — open since the 1990s. A quotient lemma reduces the question, for any alphabet, to which integer relations its numbers satisfy, turning statements about infinitely many alphabets into single finite computations. It re-derives Freedman’s published bound (L ≤ 60 whenever one number is the sum of two others) in three lines and shows it is attained; a sweep gives the first exact answers for 51 small four-number alphabets. Of eleven relations tested, only Freedman’s keeps lists short — the other ten run past 3000 letters with no end found.
Still open Proved: L=7 for every 3-letter alphabet Freedman’s L≤60 re-derived & attained First exact table, 51 alphabets 1 of 11 relations closes Chowla’s conjecture: the width of the parity barrier Jul 29 – Aug 1, 2026Count the prime factors of each integer; even or odd is a coin flip. Chowla’s conjecture says the coin has no memory, and nobody can prove even the two-flip case. An exact 2-adic descent identity reduces the conjecture to its odd core and forces a cross-scale echo of 1/√2, measured at +0.7085 ± 0.0046; a certified census to 1012 reproduces every published anchor and puts the gap between what is true and what is provable at a factor of about 5·106. Includes an audit of a preprint claiming the theorem. From the August 1 update: an independent clean-room reimplementation (different language, different machine, no shared code) reproduces all thirty prior coverage checkpoints exactly and extends coverage to every pattern of length 33; a new first-occurrence-spectrum instrument compares all sixteen million length-24 patterns against a fair coin’s exact prediction and finds nothing but noise.
Still open Certified to 1012 All patterns ≤ 33 occur Clean-room audit: 30/30 exact Echo 0.7085 vs 1/√2 Gap ≈ 5·106 Hunting the Erdős–Gyárfás counterexample July 30, 2026Join dots so every dot touches at least three lines: must some loop have length exactly a power of two? Erdős and Gyárfás expected no, and a single explicit graph would settle it. This session proves any counterexample needs at least 19 vertices (871,223,563 graphs exhausted; the reported prior bound was 17), re-derives Markström’s order-24 census clean-room (the annealer rediscovered his graph on its own), and pushes the distance to a counterexample, measured in unavoidable 16-cycles, from an exact 207 at 24 vertices down to 37 at 58, in the first window where a cubic counterexample can exist at all.
Still open Certified ≥ 19 vertices Exactly 4 survivors at n = 24 C₁₆ records ≤ 56 and ≤ 37 Hunt window 54–62 Gilbreath's conjecture — the fine structure of the Chase–Hunter–Tao constants July 28, 2026Iterate absolute differences on the primes and every row starts with 1. In the random model Chase, Hunter and Tao published this July, the expected depth-i entry turns out to depend not on the size of i but on its binary digits — so their tentative 1/i guess cannot hold in the accessible range. Includes three new exact constants, a strict sharpening of their lower bound, and — from the July 29 update — thirteen proved lenses that exclude most of the dangerous prime patterns their criterion needs, an impossibility theorem showing no fixed-order statistical axioms can prove the conjecture, and a new open problem marking the frontier.
Still open Exact c₄–c₆ 13 proved lenses Impossibility theorem R2 Open problem R3 Verified to 455,052,510 rowsTwenty-six notes so far. This page grows as they do.