Timeline
Events in this research, newest first. Each date comes from a record made at the time of the event.
Terms used on this page
- OEIS:
- the On-Line Encyclopedia of Integer Sequences, the public database mathematicians use for integer sequences. Each sequence has an A-number, such as A398957.
- Erdős problems:
- a public database of open problems posed by the mathematician Paul Erdős and others.
- Polyform, polyhex, benzenoid:
- shapes made by joining identical cells, edge to edge on a flat grid or face to face in space. Polyhexes are made of hexagons; benzenoids are molecules built from fused hexagonal rings.
- Josehedron:
- a twelve-faced solid that fills space with copies of itself.
- Octal games:
- two-player take-away games played with heaps of counters. Each heap size has a value, and a few "rare" values break the usual pattern.
- Fibonacci cube and dominating set:
- a Fibonacci cube is a network of points and links; a dominating set is a group of points such that every point is in the group or linked to one in it.
- Certified:
- in the space-filling search, a cell passed the exact checks of tiling, symmetry and counts. It does not mean the shape is new.
- Error-correcting code:
- a set of strings of symbols chosen so that any two differ in many places, so a few errors can be spotted and fixed. A record is the largest difference that can be guaranteed for a given length and number of strings.
- Ramsey number:
- the smallest number of points at which, however the links between them are colored with a given number of colors, some group of a set size ends up linked all in one color. A record attempt looks for a coloring of more points that avoids this.
- Zenodo and DOI:
- Zenodo is a public archive; a DOI is the permanent identifier it gives each deposit so it can be cited.
2026
- Correction
- Milestone
- Milestone
- Milestone
- Milestone
- Milestone
- Milestone
- Milestone
- Accepted
A401255 through A401258 published in the OEIS
They count shapes built from two more space-filling cells, an 11-faced one and a 15-faced one. A401257 and A401258 were published on October 19 and A401255 and A401256 on October 20; an outside contributor had computed the counts independently and they matched. The count of OEIS entries on this site moved from 39 to 43.
- Accepted
A400457 through A400461 published in the OEIS, all five on the same evening
They count shapes built from the cells of two space-filling solids. The count of OEIS entries on this site moved from 34 to 39.
- Milestone
Both workshop papers posted on Zenodo, free to read
The final versions of "Rediscover Before You Discover" and "How Far a Wrong Claim Travels Before It Is Caught" are now open access, each with a permanent DOI: 10.5281/zenodo.23154795 and 10.5281/zenodo.23154707.
- Milestone
Registered to present both workshop posters at NeurIPS 2026, at the Atlanta site
The conference confirmed both presentations and the registration in its invitation letter. NeurIPS 2026 runs across three sites, Atlanta, Sydney and Paris, from December 6 to 13.
- Milestone
- Milestone
- Accepted
The paper looks at what the program’s checks caught, which check caught each thing, and how early. The workshop meets in Atlanta on December 12 and 13.
- Accepted
The paper argues that an AI discovery pipeline should first re-find a recent human discovery, with almost nothing given to it, before it is allowed to claim anything new. The workshop meets in Atlanta on December 12.
- Correction
A wider search found the Josehedron’s shape type listed, unnamed, in a 2016 survey; the sequences counting its shapes are still new
The earlier check had compared it with only one table of that survey. Exact recomputation confirmed the match. The three OEIS sequences counting Josehedron shapes are still new.
- Sent out
Sent five new shape-counting sequences to the OEIS in one evening, A400457 through A400461
They waited for the editors until October 17 and were not counted on this site until then.
- Milestone
- Milestone
- Accepted
A307185 published with its new term, 5,539,121
Approved after two weeks with four OEIS editors. The entry was already counted for an earlier term, so the count of OEIS entries on this site stays at 34.
- Stopped
Stopped a month-long computer search on the twelfth Fibonacci cube without an answer, and published the attempt
After 29 days and 22 hours on one core, the solver had not decided whether 54 points can cover the whole network. The known range, 54 to 59, did not change, and the verified set of 59 remains the result. The search, its inputs and its record were published that night and later archived on Zenodo.
- From others
Terence Tao merged the seven links between OEIS sequences and Erdős problems into his public database
Problems 308, 634, 939, 942, 943, 1140 and 1161. The second merged contribution from this program, after problem 993 in August. A merge is a merge, not an endorsement.
- Correction
- Correction
Replaced the word "certificates" with "canonical-form lists" in the A307185 entry and its archive, after an OEIS editor asked what it meant
The graph software used here calls its labeling strings certificates, but to a mathematician the word means a proof. The archive now says plainly what a reader can check from the files alone and what needs the program run again.
- Accepted
A258205 published in the OEIS up to size 32
Thirteen new terms, credited by the OEIS. Its companion sequence A258204 was published on September 8.
- Sent out
Sent the new term of A307185 (a count of phylogenetic network shapes) to the OEIS
An OEIS editor worked on it the next day, and it then waited in the editors’ queue.
- Milestone
- Milestone
Published two new public code-and-data repositories, each archived on Zenodo
- Correction
Withdrew the claim that the space-filling catalog was the only machine-readable list of its kind
The claim was made before a full search. A 2026 database by Ed Pegg appears to cover the same 2016 survey and is now cited as related work. The catalog is described only as not matched against the records checked as of a given date.
- Sent out
Proposed seven more links between OEIS sequences and Erdős problems in the public database
One-line links for problems 308, 634, 939, 942, 943, 1140 and 1161.
- Stopped
Both record attempts stopped under their own rules, without beating either record
The Ramsey search reached 276 against a known bound of 381. The code search reached 49, equal to the known best, and ruled out 50 within its search.
- Result
The second stage of the space-filling search certified 165 more cells (14 from the tetragonal crystal family, 151 from the hexagonal family); none named, none matched against the records checked
Before trusting anything, each batch reproduced every printed row of a 2016 survey for its crystal family: 1,476 tetragonal rows and 958 trigonal and hexagonal rows, the latter typed in a second time by a second reader who had not seen the first copy, with no discrepancies.
- Result
Sorted all 165 certified cells by what a small nudge does to their shape type, under rules written in advance: 115 keep their type (open), 41 sit on a boundary where it changes on both sides (wall), 9 change it on one side only (one-sided)
Each cell was nudged a little in every direction tested and its shape type recomputed. Open means the type survived every nudge; wall means that in some direction it changed whichever way the cell was nudged; one-sided means it changed in one direction but not the opposite one. The quick labels guessed earlier disagreed with the computed answer on 35 of 149 cells, in both directions, and were dropped.
- Milestone
Built a machine-readable catalog of 1,583 exactly confirmed space-filling cell types, 177 of them certified, and matched it against all 3,315 printed rows of the 2016 survey
The three hand-entered tables add up to the survey’s own stated totals, and a separate recount sharing no code found zero failures. No DOI yet.
- Result
Computed two polyhex counting sequences, which had stood at size 19 since July, up to size 32, four independent ways
A258205(32) = 808,490 and A258204(32) = 750,157,868,462. The method: count only the mirror-symmetric shapes, then get the rest by subtraction from a published column.
- Correction
Confirmed the real cause of the July polyhex discrepancy, from the project’s own notes
Two programs had agreed on the same wrong values because they shared one step that identifies a shape by its corner points, which loses information for larger shapes. A different explanation recorded in August was withdrawn.
- Started
Launched two attempts at known records under rules written in advance: one on an error-correcting code over five symbols, one on a four-color Ramsey number
Preparation showed the first target could not be checked on this machine in under about five CPU-years, so it was switched to a neighboring one that could be. Three exact checkers agreed on 208 published codes before launch, and test runs caught three bugs in the search code first.
- From others
Flammenkamp updated his octal tables, taking all twelve games from July further; the July credit stays on his page
So every octal record on this site is dated July 2026, and none is described as current.
- Correction
Corrected the running OEIS count from 34 to 33 after a recount found one entry counted twice
A202182 had been counted as both an extension and a correction. This site had already shown 33 since August 29.
- Result
Computed a new term, 5,539,121, for sequence A307185, with two separately written programs agreeing exactly; not yet sent to the OEIS
Both programs first reproduced every published smaller term, down to the individual building blocks.
- Correction
Caught a wrong method name in the benzenoid census paper by checking it against the code
The paper described a backup method for larger cases, but no case was large enough to need it, and the paper had given it the wrong name. That was confirmed against the code two days later, and the paper was rebuilt around how its measures behave as molecules grow.
- Milestone
A search for new space-filling solids ended with seven cells that passed every exact check, two of them named the Satchelhedron and the Ordenhedron
None of the seven matched any shape in the reference records checked as of that date. Passing the checks means each one tiles space as claimed; it does not by itself mean the shape is new.
- Accepted
A398957 published, completing the set of Josehedron counting sequences in the OEIS
The editor asked for the description of the process to come out and the checked counting program to be uploaded as a file; both were done before publication.
- From others
The census authors asked to see the write-up, and it went to them with the section on errors in their tables marked for comment
- Milestone
Archived the code and data for five projects on Zenodo, each with a permanent DOI
These archives are the public record behind the papers.
- From others
Sent a courtesy note to the authors of the benzenoid census about the errors found in their tables
- Correction
Found a sixth error in the published census being extended, this time by reading rather than computing
Five errors had already turned up when the published tables were recomputed exactly. The sixth is a place where the paper’s text contradicts its own table.
- Accepted
One sequence counts a shape and its mirror image as different; the other counts them as the same.
- Result
Extended the published census of benzenoid molecules to two new sizes, with totals matching the published sequence exactly
The run could not report any table until its totals matched the published values in OEIS A018190, and both did.
- Milestone
Surveyed the field and prepared fifteen possible problems, checking each one against its sources
Nothing was computed: only notes, checked citations, known values to test against, and stopping rules written in advance. Going back to the raw sources caught three tables of numbers an AI agent had invented.
- Accepted
Terence Tao merged a contribution into his public database of Erdős problems
It links problem 993 to its OEIS sequence and cites the check of every tree up to 32 vertices. A merged contribution to a database, not an endorsement of a proof.
- Accepted
The last three of eleven octal-game sequences were published on the same day, completing that work
- Decision
Set aside a well-known open problem after finding others already far ahead
Earlier work had already searched far past anything a simple approach could reach, one claimed proof had been publicly refuted, and four people were listed as working on it. No computing time was spent.
- Result
The tree check finished at the next size and matched the published count exactly
No counterexample in any of sixteen separately run pieces.
- Sent out
Proposed a link between one problem in the public database of Erdős problems and its OEIS sequence
The counts were checked three times against the published data, and the proposal disclosed AI assistance as that collection asks.
- Result
Extended a Sidon-set sequence (sets of numbers whose pairwise sums are all different) and checked every value against the published terms
No mismatches with the published terms, and the underlying values were checked a second, independent way.
- Decision
Started a dated log of the program in which lines are added but never edited
When something changes, a new line records it. Dead ends stay listed, marked as dead ends.
- Result
Finished checking every tree up to a new size for an open Erdős problem, and posted the results
Another contributor had reached a smaller size in April 2026; that work is credited alongside.
- From others
The extended octal data appeared on Flammenkamp’s reference page, with credit by name
He had already reviewed the program and reran parts of the computation with his own programs. His page lists eleven games; the work covered twelve, and statements about the page use his count.
- Milestone
- Correction
A review of two and a half days of work found a discovery that an earlier summary of the results had missed
One new rare value in a game whose run had been recorded as finding none.
- Correction
The same review found that an overnight check against the OEIS had missed matches, because sequences can start counting at different positions
All ten games turned out to match existing entries, so the work was extending sequences rather than creating new ones. This was corrected before anything was sent.
- Milestone
- Result
- Stopped
A deliberately skeptical check of the WSJF paper confirmed that its central theorem was already known, and the paper was set aside
Its central result was a theorem from 1956.
- From others
Flammenkamp checked the octal program himself and confirmed the results
- Correction
Corrected a dropped digit that had sat in a published sequence for fourteen years
The correct value was confirmed four independent ways before the correction was sent. The sequence’s original author replied: "great, thanks!"
- From others
The OEIS board updated its policy on AI so that it no longer depends on which tools a contributor uses
- Correction
Caught a stray note in a draft before it was sent to the OEIS
It never reached the database. Reading the actual text before sending, rather than trusting it, is what caught it.
- Milestone
Gathered the review records from eight projects into one archive
The record now lives in one place instead of being scattered across projects.
- From others
An OEIS editor asked for three data-file uploads to be redone under a formatting rule, and was right
The rule: a full data file must not simply repeat the short list of terms shown on the entry. The files were withdrawn and redone.
- Accepted
Three new sequences published in the OEIS, with their full data files uploaded the same day
- From others
Sent the octal results to Achim Flammenkamp, who held the earlier record and maintains the reference tables
- Accepted
The added terms were published in the OEIS about an hour after they were sent
- Result
Computed the octal game 0.007 four times further than the record that had stood since 2021
Seven hours on one core of an Apple M3 Max laptop. This was the record in July; the tables have moved on since, as the entry explains.
- Result
Showed that copies of the sphenoid, an eleven-faced cell that fills space, can fit together in only one way when every cell is attached by the same rule, then counted the shapes built from it
I found no earlier counts, and no description in print of exactly how the cells fit together. The counts were later published as OEIS A397708.
- Correction
A check of a study of television transcripts against its original sources found errors in the source data, not in the analysis
Nine episodes had the wrong transcript attached, and a third of one series had been cut off by a spreadsheet cell limit. Comparing against the original source is what caught it.
- Correction
Re-ran the whole television-transcript analysis on the corrected data; every finding held
- Milestone
- Milestone
- Started
Started the program: look for small results that can be computed in under an hour and then checked
Three lines of work began the same day: continued exponentials (numbers built from towers of exponents), polyforms, and octal games.
- Stopped
Dropped the continued-exponential work the same day, after finding the result already published
The main result was already stated in the abstract of a 1996 paper. Checking for earlier work before computing anything is why this cost an afternoon instead of a week.
- Result
Computed four new terms of a polyform counting sequence that nobody had extended since it was created
- Result
Computed one of those terms again with a second, separately written program before sending it in
A different program in a different language gave exactly the same number.