Nine by Nine

Search status · 2 November 2026

Nine by Nine

Hunting a 9×9 asymmetric word square: nine nine-letter words across, nine more down, no entry repeated between the two. The first exhaustive answer landed on 10 Oct; the second, over a judged lexicon, landed at 22:03 on 15 Oct. Both say no. A third, over every judged word with no corpus filter at all, has been running since 16 Oct.

Result · complete

Over the curated 63,091-word lexicon, no 9×9 asymmetric square exists. Both halves of the start-word range are exhausted.

31,546 + 31,545 start words · 5.73 trillion nodes · 0 solutions · finished 05:30, 10 Oct

Clean pass · complete

Over the same lexicon plus 9,454 judged additions, also no 9×9 asymmetric square with a curated word in its top row or left column. Every start word searched under one frozen vocabulary; the judged words filled the other 16 places but were never start words, so a square whose top row and left column are both judged words lies outside this pass.

63,091 start words · 41.5 trillion nodes · 0 solutions · finished 22:03, 15 Oct

Full tier · running

The same sweep again over all 16,828 judged additions, coinages included and no corpus floor — the one arbitrary filter left in the pipeline, dropped. An empty result here supersedes the one above; a square here is a question for a human expert until its provenance is read.

started 08:21, 16 Oct · 400 chunks · 18 lanes on nuc13 · see the ledger for live figures

That is a real result, and a narrow one. It says nothing about English — only about this word list. Every row of a square has to be one of the 63,091 words the list contains, and the search tried all of them as a first row. The question that remains is whether the list is the binding constraint, and the rest of the work exists to answer it.

Four searches, not one

The same solver has now run four times over, against four different vocabularies. They are not interchangeable: the first two were never going to produce a publishable claim, the third did, and the fourth would supersede it.

SearchVocabularyStatusProgressWhat it is for
Baseline 63,091 curated words, LLM-generated additions complete 100% The control. No square in the plain lexicon.
Exploratory +10,054 judged words, frozen 7 Oct retired p0 2,220
p1 5,224
Developed the method, then stood aside for the clean pass on 10 Oct. Its vocabulary shifted mid-sweep, so its coverage claim was never sound.
Clean pass 9,454 judged words, frozen 11 Oct complete 63,091 of 63,091
complete on nuc13
The publishable sweep: one uniform pass, no oracle, every borrowed entry marked. Restarted 11 Oct on a trimmed tier, finished 22:03 on 15 Oct — see below.
Full tier 16,828 judged words, no corpus floor running 63,091 of 63,091
complete on nuc13
Drops the one arbitrary filter left in the pipeline: every accepted judged word, coinages included. Measured at eight and a half days.

Why a second, third and fourth pass at all

Because the baseline's vocabulary is the thing under suspicion. A square may well exist using words that are real but absent from a curated list — and the way to find out is a Borgmann search: when the solver runs out of words at a particular letter pattern, consult a language model about that specific wall, and keep the model out of everything else.

The judging that feeds the clean pass works the bottlenecks the search itself recorded, commonest first. Phase 1 — the productive prefixes — finished on 10 Oct, and the clean pass started that hour with its vocabulary frozen. Phase 2, the marginal prefixes, was reported against the frozen list rather than folded into it — and it finished on 15 Oct, all 4,345 prefixes resolved.

That report is the check the whole arrangement rested on, so here is its result. Phase 2 added 15,905 nine-letter entries the frozen lexicon does not have. Sorted into classes — each entry counted once, under the first class it matches — they are 7,295 words that sit in the corpus below the probability floor the tier was built with, 3,335 one letter away from a word already in the curated list (abfconded for absconded, the long-s scanning error), 3,127 coinages by the judge's own label, 1,286 whose clue names a surname or a city, 379 with the first letter doubled over a real word (aappeared, aapproval), 247 with a letter tripled, and 233 absent from all 20.96 million corpus rows. Three entries survived every one of those filters, and all three fail on inspection: an Icelandic place name, and two misspellings of ten-letter words with a letter dropped. So the frozen pass lost nothing, and it keeps running.

The result was not luck. Phase 2 judged the prefixes where the curated list offered nothing — which is to say the letter sequences English mostly does not use. Its 15,905 entries sit on 5,528 distinct three-letter prefixes, and none of those 5,528 begins any of the 63,091 curated nine-letter words; the log signs off having asked three models for words beginning wzk and wzm. The productive vocabulary was all in phase 1, which is exactly why the pass was allowed to start after it.

baseline 10 Oct · 0 squares exhausted judging · phase 1 all 1,313 prefixes done 10 Oct clean pass 9,454 words · no oracle complete 15 Oct · 41.5T nodes · 0 squares phase 2 + report 7,123 of 7,123 prefixes judged done 15 Oct feeds the tier freeze vocabulary runs alongside phase 2 done 15 Oct: 0 solid words, no re-freeze and since 16 Oct: the full tier, 16,828 words with no corpus floor — supersedes the clean pass if it too comes back empty
The clean pass begins after phase 1 rather than waiting for phase 2, because phase 1 covers the productive prefixes and phase 2's marginal set is where misspellings cluster. The final report, written when phase 2 finished on 15 Oct, says it was: of 15,905 additions, none was a solid real-but-unlisted word.

Where the added words come from

The tier now holds 16,828 accepted entries beyond the curated list — 16,397 real-but-unlisted words and 431 acknowledged coinages, drawn from 184,211 verdicts. The acceptance rate is deliberately brutal: nine of every ten candidates are rejected.

Words like jibsheets, xylophagy and tramsheds are the point of the exercise — genuine English that curated lists omit. Things like scriptdir and jibraltar are what the rejection rules exist to stop: a corpus built from web text is thick with programming identifiers and misspellings that look like compounds.

Every accepted entry carries a clue, because a clue is the test. If the only honest clue restates the word's own parts, the entry is not usable. And every one is labelled real or coinage, so a square resting on borrowed vocabulary declares exactly what it borrowed.

Does the program actually work?

A null result is only worth as much as the program that produced it. The standard test is Jeff Grant's 8×8 double word square, published in Word Ways in 1992 — a known-good answer with all sixteen entries distinct. Shorten the same solver to eight letters, give it a list containing those words, and it must find the square. If it cannot, every null result is suspect.

Here is the square in full. Read across for the eight rows, down for the eight columns; all sixteen are different words, which is what makes it a double square rather than a symmetric one.

Rows (across)Columns (down)
trattled hemerine apotomes metapore
nailings aloisias tentmate assessed
thamnata repeales amotions tetalite
tropisms limonias energate desessed

The 8-letter lexicon

The search draws on three tiers, each a superset of the last:

Tier8-letter wordsGrant words presentCharacter
Curated59,2588 of 16The standard the 9×9 baseline used.
Extended699,02413 of 16Still cannot hold the record.
Broad1,776,29416 of 16Complete, but thick with corpus noise.
CheckResultWhat it rules out
Square is genuine verified All eight published columns were re-derived from the eight rows, letter by letter. We are testing against a real square, not a typo.
Solver finds it found Given the 16 words at 8×8, the current solver found the square — and in both orientations, since a double square and its transpose are both valid. A search that silently missed one would be missing half the solution space.
Provenance reporting correct Reported 16/16 from the base list, 0 borrowed — the audit trail is accurate on a square whose composition we know exactly.
Finds it at realistic scale exact The sterner test, since aggressive pruning could wrongly cut the branch: a 59,266-word curated trie containing the square, searched from the start word trattled. It returned exactly one solution — the Grant square, as published. No miss, and no false positives alongside it.
Words are in the lists 8 of 16 The finding that matters most, and it is not about the code. See below.

So the program is sound. The vocabulary is a different story, and testing it against the same square produced the most consequential result on this page.

Finding · the list is the binding constraint

The curated 8-letter list is missing 8 of the Grant square's 16 words. Run at eight letters against the tier the baseline used, this search could not have found the known record — not for want of a correct program, but because the words were absent.

missing from the curated tier: hemerine · aloisias · thamnata · repeales · tetalite · limonias · energate · desessed

Widening the tier helps but does not close the gap: the next tier up, at 699,024 words, still lacks aloisias, thamnata and energate. Only the broadest list contains all sixteen — and that list is thick with the corpus noise the judging rules exist to reject.

Six of them are in the lexicon only because we put them there

Worth stating plainly, because it bears on how much the test proves. Checked against the frequency corpus and the system dictionary, six of Grant's sixteen words have no corpus occurrences at all and no dictionary entry:

trattled · hemerine · aloisias · tetalite · limonias · energate

Nothing in the ordinary evidence base supports them. They are present because the project's own record says ensuring Grant's vocabulary was in the list was deliberate — they rest on specialist attestation in the logology literature, not on measurable usage. Three more are barely above nothing: thamnata at 148 corpus occurrences, desessed at 179, repeales at 506. Only two of the sixteen, assessed and tropisms, are in a standard dictionary.

That cuts both ways, and both ways are worth saying. It means the test does not show the lexicon would independently have discovered the Grant square — the vocabulary was stocked for it. What it does show is that the solver is correct, that the list standard deliberately admits attested obscurities, and that this standard is the bar a 9×9 find should be judged against. It also measures the gap honestly: six words that a corpus-driven pipeline would never surface had to be placed by hand for a known record to be reachable at all.

This is the clearest available evidence that the baseline's null result says more about the word list than about English. The missing words are precisely the category a Borgmann search is built to recover: obscure, attested, and absent from curated lists. Thamnata, desessed and energate are not noise — they are a published record's vocabulary, and the lexicon the baseline searched did not contain them.

It also cuts against the trim described below. The clean pass now requires every borrowed word to score at least 0.7 on a single corpus field, english_probability; a word with no corpus row scores nothing at all, so this is exactly the filter that would have discarded those six. The trade is deliberate — a sweep that will not finish produces nothing, whatever its lexicon — but it means this pass is blindest to the one category of word that made the known record possible. The 7,374 words it drops are still in the verdict cache, and they are the natural input to a later targeted search rather than a full sweep. That filter has since been measured against an expert's scores and turns out to be a far worse selector than it looked — What the 0.7 filter actually selects, below.

What would count as a find

A found square prints its own audit trail: eighteen entries, with each one marked base, fill or oracle. A result reading provenance: 15/18 from the base list, 3 borrowed is a different claim from one reading 18 of 18, and the difference is exactly what an expert reviewer needs.

What is not yet settled

The exploratory run's coverage cannot be quoted

For about a day, four solver processes shared one checkpoint base — two writing each progress file — because a reboot entry and a manual start both slipped past the guard. The duplicates are gone and a lock now makes a second instance impossible, but that run's "searched everything up to N" figure is not trustworthy. Its value was developing the method, and the clean pass supersedes it.

A tier is not free vocabulary — it is a multiplier on search time

The first attempt at the clean pass managed about 5 start words per lane-hour against the baseline's 30 to 34 — nine months to a year for the sweep. This page said four to seven weeks, which was wrong. Rather than guess at the cause, we measured it: identical start words, 300 seconds per configuration, one variable at a time.

Fill words addedStart words / 300sReading
010The baseline lexicon alone.
89510A small tier is free.
9,4535Half the throughput.
15,1571The cliff.
16,8281What was running.

Two separate costs, and both are real. Vocabulary is superlinear: 895 extra words cost nothing, 15,000 cost a factor of ten, because every added word lets more partial squares survive pruning at every cell. And the oracle costs 5× on its own — at 9,453 fill words, attaching it dropped throughput from 5 start words per 300s to 1. Even answering from cache, every dead-end prefix is a socket round-trip through a single-threaded server.

So the pass was restarted on 11 Oct with both costs cut. The tier is now 9,454 words — accept-verdict, real-only, and scoring at least 0.7 on the corpus's english_probability field — and the oracle is gone entirely. Dropping it costs almost no vocabulary, since the fill list is built from the same verdict cache the oracle would have answered from, and it makes the result cleaner to defend: the lexicon for this sweep is one enumerable file, not a live service whose answers could drift mid-run.

What the 0.7 filter actually selects

An earlier version of this page described that trim as keeping the words "with real corpus support". That was the wrong description, and it is worth replacing with what the field is rather than what we hoped it was. english_probability is one column in a 21-million-row frequency database built from web text. It was used as the gate because it was there and because thresholding it cut the tier to a size that finishes. It is not a judgement of wordhood, and that is now measured rather than suspected.

On 13 Oct the first batch of calibrated scores arrived from Eric Chaikin, the logology expert on this project: 167 strings, each scored 0–100 for whether a logologist would accept it. 125 of them have a corpus row. The rank correlation between his score and english_probability is +0.05 — no relationship at all. Of the 50 strings he accepts, the 0.7 floor keeps none: forty score below it and ten have no corpus row at all.

A sample of the disagreement. The expert column is his 0–100 score where he scored the word, and — where the word is ours rather than his.

Dropped by the 0.7 floorepexpertAdmitted by itepexpert
qwerty0.6488scriptdir0.84—
werewolf0.6786parvclass0.755
serendipity0.6683chavtastic0.7535
bigamy0.6579motorcamp0.84—
discombobulate0.5076mobilefun0.75—
kerfuffle0.5874mosschops0.75—
schadenfreude0.5972motofocus0.81—

The right-hand column is not a hypothetical. Every one of the 9,454 words in the tier the sweep is running on today scores 0.70 or higher by construction, so every one of them outranks serendipity on the signal that chose them. Here is a consecutive stretch straight out of the live file: morticias, moseshair, mosschops, motichoor, motocaddy, motofocus, motomania, motoracer, motorball, motorcamp, motorcare. And scriptdir — named above as exactly what the rejection rules exist to stop — sits at 0.84, higher than almost anything a dictionary would list. The judging caught it; the trim would not have.

The reason is structural. A well-formed compound of common English parts scores high on a model of what English text looks like, however invented the compound is, while a loanword, a coinage-turned-standard or an idiom scores low. So thresholding this field selects for the web-compound class the judging prompt was written to reject, and against the attested obscurity a Borgmann search exists to recover. Both of Eric's scored phrases, kick the bucket and in the long run, score zero. A correction to how that was put here before: this page said they had no corpus row whatsoever, and that is wrong. The database holds 185,466 rows containing a space, both phrases among them — they carry a row with english_probability and count both zero. Common phrases behave the same way: new york and ice cream are present and also score zero. So spaced phrases are indexed but unscored, which leaves the conclusion intact — the 0.7 floor drops them either way — while making the reason the opposite of what was claimed. It also means corpus frequency cannot rank a de-spaced phrase at all, which matters because de-spacing a phrase is one of the standard routes to a record entry.

What this does and does not change. It does not change the speedup: the trim bought the factor of ten it was meant to buy, the sweep now finishes in weeks instead of a year, and that is why it is still running. It does change the claim that can be made about the tier — these are not the best 9,454 of the 16,828, they are the 9,454 one frequency field happened to rank highly, and the honest reading of a null result over them is correspondingly weaker. The fix is not a different threshold on the same field: it is to use the calibrated scorer as the gate at the next freeze, once there are enough scored points to fit one. Not mid-sweep — a resumed pass must never change the lexicon it started with, and 50 reviewed points is too thin to refit a gate on.

The filter, tested against the one square that is known to exist. Jeff Grant's 1992 8×8 double square is sixteen real entries that a logologist accepted and Word Ways published. Looked up in the same frequency database, their english_probability runs from 0.2813 to 0.7975, and the 0.7 floor would have admitted exactly one of the sixteen — assessed. The record's whole vocabulary, bar that one ordinary participle, is what this gate is built to throw away.

Grant's 16 words by english_probability, with corpus count assessed 0.7975 58,924,304 <- the only one the 0.7 floor keeps trattled 0.6800 2 hemerine 0.6800 2 tetalite 0.6400 1 limonias 0.6400 1 energate 0.6400 1 tropisms 0.5796 142,975 aloisias 0.5600 3 tentmate 0.4827 9,814 nailings 0.4666 6,294 amotions 0.4654 6,086 metapore 0.4238 1,929 apotomes 0.4040 1,116 repeales 0.3756 509 desessed 0.2881 180 thamnata 0.2813 149

Read the top of that table against the bottom. trattled and hemerine, with two corpus occurrences each, outscore apotomes and metapore, which have a thousand and two thousand — because they are pronounceable English-looking strings, and the field measures exactly that. Five of the sixteen sit at a count of one or two, which is noise and may well be circular: scraped word lists are themselves in the corpus, so a record word can appear in it because someone published the record. Frequency cannot distinguish the vocabulary of a record from a typo, and it ranks a well-formed invention above an attested rarity.

The same square answers a question about the judge rather than the filter. Asked to rule on those sixteen words from knowledge alone, this model recognises six without hesitation — assessed, tropisms, nailings, tentmate, apotomes, amotions — accepts metapore as specialist anatomy, treats repeales as an archaic spelling, reads aloisias, limonias and thamnata as proper names rather than words, and does not know trattled, hemerine, tetalite, energate or desessed at all. Six of sixteen, seven generously. And the six it accepts are precisely the six with real corpus counts, from 1,116 up to 58.9 million, while the five it rejects outright are the five at one or two occurrences. The judge is not an independent witness to the language; it is reporting the same frequency distribution the filter uses. A Borgmann search gated on this model's recognition alone would have rejected most of the vocabulary that makes the known record possible — which is the whole case for a human expert's ruling, stated on a square that exists rather than on one that might.

The expert's answer, and the target this project is actually aiming at

That table went to Eric Chaikin on 21 October with one question attached: would you accept all sixteen of these in a grid today? He answered on 22 October, and the answer changes the bar rather than the verdict. He replaced his 0–100 scale with a 0–5 grade across two targets — Crosswords, and Wordplay, which he says explicitly includes generic word squares which are not otherwise crossword grids. A 9×9 double square is Wordplay. So every earlier reference on this page to his accept line of 70 was quoting his crossword bar, and on the target that applies here the line sits at 3, which he maps to .60.

grade meaning maps to 5 Crossword desirable (sparkling fill) >= .90 4 Crossword OK (good fill) >= .70 3 Crossword avoid but use if needed / WORDPLAY OK >= .60 2 Crossword no / Wordplay avoid but use if needed >= .50 1 both low quality, last resort only > 0 0 nothing, do not use 0 his grade for each of Grant's sixteen, against the field that was gating them Eric ep corpus count assessed 5 0.7975 58,924,304 trattled 4 0.6800 2 hemerine 3 0.6800 2 tetalite 1 0.6400 1 limonias 3 0.6400 1 energate 2 0.6400 1 tropisms 5 0.5796 142,975 aloisias 3 0.5600 3 tentmate 1 0.4827 9,814 nailings 2 0.4666 6,294 amotions 3 0.4654 6,086 metapore 2 0.4238 1,929 apotomes 3 0.4040 1,116 repeales 2 0.3756 509 desessed 2 0.2881 180 thamnata 2 0.2813 149 (revised up from 0, see below)

Read by target, that is the sharpest statement this page can make about what standard a record attempt has to meet. As crossword fill, only three of the sixteen are good fill at all — ASSESSED, TROPISMS and TRATTLED — and the remaining thirteen are avoid, no or worse. As wordplay, eight of sixteen are outright OK, fourteen are usable if needed, and — once the single revision described below is applied — all sixteen are usable at least as a last resort, with none unusable for anything. A published record is therefore near-unusable in a crossword and fully usable in a word square — the target distinction validated against a square that exists rather than against a hypothetical one, which is the only way this project has found to test any of these rules.

What grade each entry must reach, in his words (23 October). Asked directly whether a record 9×9 needs every one of its eighteen entries at 3, or whether 2 will do, Eric answered that it is neither: the bar ratchets. Interest Level depends on degree of difficulty - initially, we would accept even down to Score 1. Once we have otherwise high quality results including a 1, then tighten to 2, etc. Ideally all >=3 would be a satisfying result. So a first 9×9 counts if no entry is a 0, each later find has to beat the one before, and all-3s is the satisfying end state rather than the price of entry. On his grades, all sixteen of Grant's entries clear that first bar. He added that toponyms and their legitimate variant forms are acceptable for wordplay exercises such as word squares — which points the GeoNames rule toward admitting more, not less. What the ratchet does not settle is the 430 coinages in the full tier: he has not graded one, and whether an invented string can earn even a 1 is the question a coinage-bearing find would have to put to him.

One conversion this page will not make. His 60-69 band being Wordplay-OK does not promote the 28 new place names that toposcore happened to put in 60-69, because that band is on his scale and those scores are on ours. The nine anchors further down measure toposcore running about 38 points low with a rank correlation of −0.16 against him, so the two do not share an axis, and sliding one onto the other is precisely the error that cost this page its +35 named-entity offset.

The one 0 he first gave was thamnata, and he was explicit that it was a statement about what he could find rather than about the word: I simply could not find any reference for THAMNATA… that is the level of attestation I could actually find - which was 0. Asked whether that makes his standard tighter than Grant's, he said in this case I suppose yes. It no longer stands at 0 — he revised it to 2 the same day, and the reason he gave matters more than the grade. That exchange is below, under the gazetteer.

How fast it actually goes, measured twice

The solver runs at the lowest priority by design and yields to everything else on the box, so the first guess was that throughput is a function of CPU share. Three measurements later that turns out to be only half true.

Starved. For its first 17 hours the two lanes shared four cores with unrelated number-theory jobs and held about 28% of a core each. From the checkpoints: 230 start words for lane 0 and 444 for lane 1, or 13 and 26 per lane-hour.

A 90-minute probe after a 12 Oct reboot cleared those jobs, sampling the checkpoints every 15 minutes and recording each lane's CPU share beside every sample: 47 and 40 per lane-hour at around 70% of a core each.

13:04 p0=420 p1=32440 load=8.8 cpu 13:19 p0=430 p1=32460 +10 +20 load=12.2 76% 76% 13:34 p0=450 p1=32470 +20 +10 load=6.3 74% 73% 13:49 p0=460 p1=32480 +10 +10 load=6.4 90% 90% 14:04 p0=460 p1=32490 +0 +10 load=10.4 73% 73% 14:19 p0=470 p1=32490 +10 +0 load=11.6 34% 35% 14:34 p0=490 p1=32500 +20 +10 load=5.3 68% 66%

The 16.6 hours since that reboot, at the same average CPU share of 71%: 59 and 117 per lane-hour, and taking the overnight stretch alone, 67 and 139. Lane 1 is now three and a half times faster than the probe measured it, on the same fraction of the same machine.

The sweep moved to nuc13 on 15 October — 20 threads and nothing else to do, against four shared cores here — and finished there the same day. Its 18 lanes took 21.8 hours to clear all 200 chunks, averaging 1873 start words per hour between them against 298 on nuc5. On the four cores it started on, the same work would have run into November.

Which overturns the first explanation. Between the probe and the overnight window the CPU share did not change — both average about 70% of a core — and the throughput changed by up to a factor of three. What dominates is which start words a lane happens to be working through: a nine-letter first row that admits many second rows opens a tree orders of magnitude larger than one that strangles at the second cell, and those words are not evenly distributed through an alphabetically sorted list. CPU share matters, as the starved figures show, but inside a well-fed run the lexicon's shape matters more.

So the schedule was a range rather than a date — and the range closed at 22:03 on 15 October, eleven and a half hours after the move, the sweep complete — nothing left to search, 100.0% of the range swept — against the nine months the untrimmed configuration implied. One bias in the figures above is worth keeping, because it was systematic rather than noise: every reading taken while the sweep ran understated the finish, since the work was handed out hardest-first — the expensive early-alphabet chunks ran while all eighteen lanes were busy, and the cheap tail went in minutes. Progress was checkpointed every 10 start words, so each sample was quantised to that step.

On 15 October the sweep moved to a bigger machine, and the schedule changed from days to hours. For its first three days the clean pass ran on a four-core desktop that also carries a browser, a cron'd mathematics job and its owner's work. Both lanes ran at nice 19, which makes the search the first work on the machine to yield, and they were getting 85% of a core each while the load average sat between six and ten. Measuring where that load came from is worth recording, because the intuitive answer is wrong: the two lanes were 2.00 of a load average of 10 and about half the CPU in use. A load average counts threads that want to run, not work getting done, so a lane that is always ready contributes a whole point to the figure even when competition has squeezed it to a third of a core. The largest single contributor was not a big program at all but a shell loop in the other project that re-scanned an 81 MB file once per item — 551 short-lived processes inside a two-minute window. Stopping the two lanes confirmed the arithmetic: CPU in use on that machine fell from 4.0 of 4 cores to 2.2, the mean count of threads wanting to run from 7.1 to 3.6, and what is left is still mostly that shell loop.

The search now runs on a 20-thread box that was sitting idle, and this is the same search rather than a new one, which matters for the coverage claim. The solver was rebuilt there from the same word_square_correct.cpp with the same compiler and is byte-identical to the binary that had been running (md5 675fdf3be80c6a359c137c224c0aaa70); the runner re-checks the md5 of the binary and of both word lists before it will start, so the vocabulary cannot drift mid-sweep; and the ranges continue exactly where the old checkpoints stopped, so nothing is re-searched and nothing is skipped. Throughput was measured rather than assumed — sixteen lanes over five minutes covered 420 start words, about 5,040 an hour against the desktop's 298.

The work is split into 200 chunks of 206 start words fed to eighteen lanes, rather than one long lane per core. The reason is in the numbers above: the cost of a start word varies by a factor of sixteen across the range, so fixed lanes would finish in the time of the unluckiest one. Small chunks drawn from a queue let a core that gets an easy stretch come back for another, and because the queue is ordered hardest-first the slow early-alphabet work is done while every core is still busy. Nothing in the coverage depends on the move: a chunk resumes from its own checkpoint, and a re-run skips the chunks already finished.

One comparison deliberately not drawn: the baseline's 30 to 34 start words per lane-hour was measured under load conditions nobody recorded, so the fact that this pass now exceeds it does not mean an enriched lexicon searches faster than a plain one. The controlled benchmark above is the only sound statement about what vocabulary costs.

The 124 start words searched by the untrimmed attempt were discarded, which costs nothing. Its vocabulary and checkpoints are kept as a record.

The general lesson is worth more than the schedule. A judged tier is not free vocabulary — it is a multiplier on search time, so the discipline that makes a Borgmann search finishable is strict judging rather than generous judging. Every marginal word admitted is paid for twice: once in the risk that a found square rests on it, and again in the weeks it adds before any square is found at all.

Its first day was slower still: a watchdog meant to restart crashed searches could not tell a deliberate retirement from a crash, and brought the exploratory run back six minutes after it was stood down. Four solver lanes shared four cores for sixteen hours. The revive rule is gone, and the watchdog now also watches for a square found by the clean pass — which, until then, it would not have announced at all.

The judging survives a reboot less well than the search does

The 12 Oct reboot was a fair test of the recovery machinery, and it half passed. The search came back unaided: a reboot entry restarted the script, which reused its frozen fill list and resumed both lanes from their checkpoints, losing at most ten start words each. The judging did not — nothing revives it, and the same process is what writes the phase 2 report, so a silent death there would have cost the report rather than the sweep.

Restarting it needed care for a reason worth recording. The judging pass walks its queue of recorded bottlenecks commonest-first and deliberately re-resolves each one. Verdicts on words are cached, so re-judging them is free — but coinage is generation, not judgement, and nothing caches it. A plain restart would have re-asked all three models for the 1,062 coinage prefixes it had already paid for before reaching any new work: about a day of wall clock, and real spend, for information already on disk. Trimming the already-judged prefixes out of the queue first brought it back to 4,345 remaining — exactly the 7,123 it started with minus the 2,778 it had done.

The lexicon is the open question — and now a measured one

The Grant test settles the direction of doubt. A curated tier missing half a known record's vocabulary is not a tier that can rule anything out, so the baseline's null result should be read as a statement about the list, not about English.

The clean pass has come back empty. All 63,091 start words, 41.5 trillion nodes, no square — so the conclusion is the one the paragraph above anticipated: no 9×9 asymmetric square exists over this lexicon plus the 9,454 judged additions that one corpus field ranked highly, among squares whose top row or left column is a curated word (the judged words were fills, never start words). That is a stronger claim than the baseline's, still not a claim about the language, and weaker than the 16,828-word version would have been had it been finishable. Widening further means more judging, and judging is bounded by an expert's time rather than by compute.

And now the pass that was unaffordable

One arbitrary step survived in the pipeline that produced the result above: the 0.7 corpus-probability floor that cut the judged tier from 16,828 words to 9,454. The section further up shows that floor does not measure word quality — against expert scores it does not correlate at all, and it selects well-formed web compounds over exactly the attested obscurity a Borgmann search exists to recover. It was imposed for speed, because 16,828 words ran about ten times slower and four cores turned nine days into nine months. With the big machine now idle, that reason is gone, so the full tier is being swept: 16,828 entries, 16,397 real and 431 coinages, every accepted judged addition with no corpus filter at all, and 1,487 of them carrying a marginal verdict rather than an accept.

What that tier does and does not contain is worth stating exactly, because it decides what a hit would be worth. Of the 16,828 entries, 1,487 carry a marginal verdict rather than an accept — admitted here, excluded from the pass above. Measured against the corpus: 615 of them are real words with support at or above the old floor (grayfrier at 0.97, girlguide and battlebus at 0.94, hattricks at 0.90), 476 are real words below it, and 391 are coinages with no corpus row at all. Not one is a real word with no corpus row — and that is precisely the Grant class, since six of Grant's sixteen words have no corpus occurrences and no dictionary entry. A corpus-fed pipeline cannot produce a TRATTLED; what fills that slot here is coinage, and a coinage is not evidence. That gap is what an expert word list closes and nothing else in this pipeline can.

Asked to rule on a sample of the marginal entries, Eric Chaikin's first move was to correct the question — whether we are after valid entries rather than solid words — and on that reading he put hattricks, ladiesman, golfwidow and muslimdom at 70 or above, the rest needing a look-up. That distinction is the one that matters for a record attempt: a dictionary-solidity standard makes the known 8×8 unreachable, so validity is the right bar, and the bar itself is worth more than any individual ruling.

Asked next for one rule per class instead of 1,487 verdicts, he gave four, and they are stricter than the guess that preceded them. Archaic spellings out (publisshe, inquyring, culloured). Foreign-language words out unless naturalised — his test is the word appearing uncapitalized nonitalicized in English — which clears 14 of 366. Dialect and slang out. And toponyms in, if the name appears anywhere in GeoNames, with no city-or-country restriction. Applied mechanically to all 1,487 by the clue the judging model itself wrote, that refuses 547 outright and admits 14. The prediction sent to him with the question — archaic in, dialect in, foreign out — would have admitted 195 words he refuses, all in the same direction: the model judging this list is too generous, and an expert's rule is the only thing that showed it.

One of his answers is a result about the corpus rather than about the words. Of the six plain compounds he ruled on individually, the three he refused carry corpus probabilities of 0.97, 0.94 and 0.93; the three he admitted carry 0.86, 0.83 and 0.80. The separation is perfect and it runs backwards — every word he threw out scores above every word he kept. With six words that ordering arises by chance one time in twenty, so it is suggestive rather than settled, but it points the same way as the +0.05 correlation above and it sharpens the claim: on well-formed compounds a high english_probability is weak evidence against validity, because the field measures how much a string looks like English text and an invented compound of common parts looks exactly like English text. The 913 compounds he has not ruled on cannot be sorted by it either — 243 sit in or above the band where his own six disagree with each other.

His toponym ruling opened a class this search had never looked at. Candidates here come from a corpus scan for common nouns, so a place name could not be proposed at all — not rejected, invisible. Prompted directly for nine-letter place names, the model offered 116; 75 were already in the curated 63,091, so the list is not short of toponyms, two more sit in a judged tier, and 39 are new: HIMALAYAS, PATAGONIA, BYZANTIUM, KATHMANDU, RAJASTHAN, GUANGZHOU, ISLAMABAD, ANDALUSIA, AQUITAINE, PHOENICIA, CHERBOURG, POMERANIA, THURINGIA, WALLACHIA and the like. Two things about them matter more than the count. Every one sits below the retired 0.7 corpus floor — the highest is 0.69 — so that filter would have deleted the entire class. And 36 of the 39 sit on three-letter prefixes that already begin curated words, so a partial square can actually reach them, which is exactly what phase 2's 15,905 additions could not do on any of their 5,528 prefixes. Reachability, not plausibility, is what separates vocabulary from decoration.

The ruling also changes what kind of test this class needs. For every other class an expert's scale is the instrument; for toponyms his answer is all fine if anywhere in GeoNames, which replaces a judgement with a lookup in a gazetteer — an external, checkable, repeatable gate that does not depend on a model's opinion or on an expert's time. That is the better outcome, and it is not yet run: no gazetteer is held locally, so the 39 names and the 9 toponym-clued marginals that survive a corruption check (4 others turned out to be our own generator's noise — dknighton for Knighton, hbcrawley for Crawley) are pending that one file. Nothing has been added to the running sweep's lexicon: it is md5-guarded and must not change mid-pass.

A correction to the paragraph above, and to what this page said earlier today. The claim that 34 of those place names clear Eric's 70 came from a +35 correction applied to the model's own scores for named entities — a correction measured over all eight named-entity entries in his list. Audited by the parallel crossword-stack project, that number does not survive: the two halves of the shared fit/held-out split put only 2 of those 8 words in the fitting half, so the offset was fitted on the very words it would then have been tested against, and a 2-word sample is below any floor worth having. Refitted honestly, the class has no defensible score correction at all — and with no correction, none of the 39 clears 70. The holdout now applies to every fitted quantity on this project rather than to the accept cut alone: bias, spread, rank correlation and each per-class offset are fitted on the fit half, and the held-out half is read by the program and never by the model, reported as counts with no word printed. On that footing the model's own cut of 92.5 reproduces Eric's 70 on 91% of the held-out half, admitting 1 word he refuses and refusing 6 he admits. The place-name class survives all of this on Eric's rule rather than on a score, which is the point: the gazetteer test needs no calibration to be trusted.

The gazetteer has now been run. GeoNames' full dump — 13.47 million rows, every name reduced to a-z as an entry would be written, 11.56 million keys — settles the class in one pass. 38 of the 39 place names are there; the single miss is MANCHUKUO, the Japanese puppet state, which has no entry. And the rule's precision can be measured rather than assumed, because the tier carries a built-in control: 431 strings our own judge labelled coinages, invented rather than found, where any hit would be pure coincidence. The gazetteer admits none of them — zero of 431. So on nine-letter strings Eric's rule separates the class it should admit from invented matter by 97 points, with no false admissions at all, and it correctly rejects all four of the corrupted strings this pipeline produced (dknighton, hbcrawley, aswidnica, cizenburg) because a typo is nobody's place name.

One caution belongs beside that, and it is the opposite of a quibble about the rule: it is a limit on where the rule may be pointed. Applied not to toponym candidates but to the 1,487 marginal entries as a rescue, the same lookup admits 34 of them — and they are coincidences rather than places. olivander is a farm in South Africa and the clue says wandmaker in Harry Potter; blackmesa is a mountain in Arizona and a fictional research facility; altovista is a district of Aruba and a defunct search engine; hitormiss is an Australian homestead. No mechanical qualifier fixes this — requiring a place's own name and a true toponym feature class still leaves 21 of the 34, while requiring a population above zero would delete HIMALAYAS, BYZANTIUM, KARNATAKA, HINDUSTAN, TURKESTAN and WALLACHIA, which are real. The conclusion is about scope: a gazetteer answers is this place real? and it answers it very well, but it cannot answer is this string an English word?, and the second question is the one the marginal class asks. Eric's rule is therefore sound for entries proposed as place names and unsound as a general amnesty — which is how it will be applied here. A related finding from the parallel crossword project sharpens the point: on three-to-five letter strings the same rule flipped 751 junk entries, because a short random string lands on some hamlet's name far more often than a nine-letter one does. The rule's precision depends on length, and this project's nine letters are where it is strongest.

The one entry no gazetteer carries

Eric's 0 for thamnata turned on attestation he could not find, so the obvious move was to point his own toponym rule at it: THAMNATA is a place name, and his rule admits a name appearing anywhere in GeoNames. The prediction was that his grade and his rule would collide. They do not. Against the full dump — 13,474,846 rows, 22 October — thamnata has zero rows, and so do thamnatha, tibneh, timnah and khirbettibneh. The only near match is timnath, which resolves to a town of 625 people in Colorado — a modern homonym, the same failure recorded above for GUANGZHOU and JALALABAD. So the gazetteer refuses the word too — which at the time agreed with his grade of 0, and then he told us the rule was the thing that was wrong.

What that establishes is a third limit on the rule, and a worse one than the two already stated. The place is real: THAMNATA is the Latin Vulgate's rendering of Thamnatha, a fortified city in Judea, identified with the ancient Timnath at modern Khirbet Tibneh. A gazetteer of 13.47 million present-day rows simply does not carry it. So besides being unable to answer is this string an English word?, and besides admitting modern coincidences, the rule under-admits the historical layer, and what it offers there instead is a present-day homonym. That also defeats the mechanical handle this project built for Eric's own instruction that toponyms have a time dimension: the historical feature codes PPLH, RGNH and ADM1H can only classify a place the dump contains, and this is a place it omits.

Then the expert overruled his own rule. Sent the zero result, Eric answered that historical toponyms wouldn't necessarily be in Geonames, and that given a real citation I'd give it a 2. Wouldn't use it in a publisher crossword but acceptable in a wordplay exercise if needed. Chris's Vulgate citation is that citation, so THAMNATA moves from 0 to 2 — and with it the last of Grant's sixteen becomes usable as wordplay.

The consequence for the record attempt is worth stating plainly, and it is not the consequence this section was written to state. Under Eric's Wordplay standard, the known 8×8 record is reconstructible: all sixteen entries are usable, eight of them outright. What is not reconstructible is the record under a GeoNames-based toponym rule, which still refuses the word and would still refuse it today. The rule and the expert who proposed it now disagree, and he is the standard — so the rule is the part this project has to fix. The binding constraint on a Borgmann search is not the judge and not the compute; it is the evidence base, and the sharpest version of that claim is this one, where the judge could see past a gap in the evidence and the mechanical rule built on it could not.

Two refinements to the paragraph above, both of which weaken it before they strengthen it. First, that control is easier than it looked. 431 invented strings cannot test the failure mode that matters most for a gazetteer — a string that is a real place and not an English entry — because an invented string almost never names a place. The crossword project's cohort is corpus text, which contains gazetteer text, and at ten letters and above it still turns up PINNISJOCH (a saddle in Austria) and INNISKILIN (a village in Jamaica). So the zero is a true measurement of one risk and silent about the other, and the honest reading of it is narrower than "no false admissions": invented matter does not sneak through, real foreign micro-toponyms can, and only the requirement that the judge proposed the entry as a place name keeps them out. Both projects have now scoped the rule that way.

Second, each of the 38 hits was checked against what its own clue claimed, because a string can be in the gazetteer for the wrong reason — and that check first appeared to fail six times. It did not. The failure was in the test: the shared index keeps one representative row per name, so a word whose representative sat in the wrong country looked like a homonym. Read against the full 13.47-million-row dump instead, every one of the six holds — jalalabad has 168 rows including the Afghan city of 271,900, patagonia has 19 including the Argentine region, phoenicia is there as a historical region of Lebanon. 38 of 39 verified against their own clue; MANCHUKUO alone has no entry in 13.47 million rows. A reduced index answers whether a string is in the gazetteer, not whether a place is, and those are different questions.

That check also found the thing that makes Eric's outstanding ruling sharper than it looked. BYZANTIUM is in GeoNames three times: a homestead in Australia, a building in Korea, and — correctly — as an alternate name for Istanbul, population 15.7 million. RAJASTHAN likewise: its primary-name row is a locality in Pakistan, while the Indian state of 68.5 million is carried as an alternate. So on a primary-names-only reading both are admitted for the wrong reason, and on an alternates-count reading both are admitted for the right one. The question is not a third of the class, as first put to him; it is a handful of entries, and for those few it is the difference between a correct admission and a coincidence that happens to agree.

He answered, and the answer dissolves the question rather than settling it. Asked whether an admitted place is a full accept or a flagged one, Eric replied with a method instead of a verdict: score toponyms by F(status, population, size), where status is a GeoNames tier — continent > country > capital city > well known city > major feature > town > village > etc > minor feature — then overlay with lexical presence and corpus frequency, expand to include multiple forms for all toponyms, adjusting scoring accordingly for esch, remember toponyms have a time dimension, and bias against source junk mainly by length: Any 15-letter place that fits a crossing is usable whereas obscure 3 letter places should only be used as a last resort.

So the gazetteer is not the gate this page called it two paragraphs ago — it is one input of five to a graded score, and the correction matters because a gate admits or refuses while a score ranks. His third point also settles the alternates question without being asked it directly: multiple forms all count, each scored on its own, so BYZANTIUM is not refused for reaching Istanbul through an alternate name — it is discounted. And his fifth point arrives independently at the length effect measured here from the other direction: the same rule that flips 30% of rejected three-letter strings flips 2.7% at nine letters.

The method is now implemented rather than paraphrased. Status comes from the feature code, population and the historical marker straight from the dump — GeoNames flags historical places with its own code suffix (PPLH, RGNH, ADM1H), so the time dimension has a mechanical handle rather than needing a judgement. Lexical presence is membership of the curated 63,091; corpus frequency comes from the 21-million-row database. One factor is not measured and is not quietly folded in: the dump carries no area field, so size is approximated from the feature code and kept in its own column. The scoring ran over all 116 candidates, and the check that matters is the 75 already in the curated list — toponyms a human word list has accepted, so a scorer that refused them would be wrong about words known to be fine:

80+ 70-79 60-69 50-59 40-49 <40 curated (75) 20 30 17 5 2 1 new (41) 1 5 24 9 0 2

64% of the 75 clear his 70, which is the result that makes the weights usable — and that figure fell from 69% when the form-selection rule was corrected a third time, because picking the right place scores lower than picking the most flattering one. The clue's country and feature type now order the candidate forms lexicographically rather than adding a few points: as a bonus it was too weak to do its job, since a clue reading Italian river gave the Lazio river only six points while a stream ranks 26 on the tier and a village 50, so LIRI went on scoring off a village in Cameroon. Ranked first and scored second, LIRI resolves to Fiume Liri in Italy and scores 2 — the right place with an honest number, where before it was the wrong place with a flattering one. The drop of five points in that validation rate is the price of that correctness, and it is the direction a measurement should move when a bug is removed. The 41 new names behave quite differently: none reaches 80, three clear 70 (HINDUSTAN, MACEDONIA, ANDALUSIA), and 28 of 41 land in 60-69 — just under the line, in the band Eric's own calibration work calls the whole ballgame. MANCHUKUO scores zero, having no entry at all. So the honest verdict on this class is neither the 34 accepts first claimed here nor a rejection: these are real places of middling notability, clustered immediately below the accept line, and that is a finding about the class rather than a failure of it.

Choosing the best-scoring form per name meant the scorer picked the most flattering matching row, and twice it picked a different place: guangzhou scored against Gwangju in Korea, jalalabad against a province in Kyrgyzstan rather than the Afghan city. Both are now fixed, by letting the judge's clue select the form while F() scores it — and the two wrong turns on the way there are worth recording, because each was a plausible design that measurably made things worse. Using clue agreement as a filter rather than a preference let a crude parser overrule the gazetteer: it read South American country as the United States and dropped ANDALUCIA from 53 to 9 by restricting it to a hotel. Adding clue agreement to the score was worse in a subtler way — nearly every clue names its own country correctly, so the bonus was a near-constant uplift that lifted the share of new names clearing 70 from 5% to 51% and quietly turned the judge wrote an accurate clue into notability. Selection and scoring have to stay separate. A third bug was older than either: accented letters were being dropped rather than folded, so Andalucía reduced to andaluca — not a near miss but a different real village in Argentina, scored in its place. And the weights themselves are reasoned, not fitted — they cannot be fitted, because the only toponym Eric has scored is skopje, north macedonia at 83, and that entry sits in the held-out half of the shared calibration split. Fitting to it would repeat precisely the error that cost this page the +35 offset. The cheap fix is a two-line ask: eight toponyms scored across his own tiers, from continent to minor feature.

That ask has acquired a second question, from a tension inside his own answer. His first point ranks a minor feature at the bottom of the tier; his fifth says any 15-letter place that fits a crossing is usable. Those collide on exactly the entries a long crossing wants: ALASKAPENINSULA scores 22 and ALABAMARIVER 17, because GeoNames files Alaska Peninsula as a CAPE and the Alabama River as a STM — both minor features with no population, and a six-point length bonus cannot lift them. Either the length term has to dominate at fifteen letters or large water and terrain features need a higher tier than a creek. It is not a weighting we should choose for him, so both projects are leaving the length anchor marked unsettled rather than quietly setting it.

He has now answered, with nine scored anchors, and they say this scorer is badly miscalibrated. On the tier question he ruled that a major anything … should outrank minor things, unless they are of a category that is "notably major", adding that Claude should apply a reasonableness standard as won't be able to enumerate all cases. Then he scored nine toponyms spanning his tiers. Against the implementation:

word Eric toposcore gap wallachia (control) 85 57.4 -27.6 alaskapeninsula 80 23.4 -56.6 alabamariver 80 19.3 -60.7 adeliecoast 80 17.8 -62.2 chippingsodbury 75 45.8 -29.2 alhudaydah 75 60.0 -15.0 lorient 75 47.0 -28.0 lomita 70 33.0 -37.0 kematen 70 42.8 -27.2 mean signed error -38.2

Eric admits nine of nine; the scorer admits none of them. Mean error −38.2 points, and the rank correlation between the two is −0.16 — so it is not merely harsh, it orders them slightly backwards, with his three highest (a peninsula, a river and a coast, all 80) scoring lowest at 18 to 23 because GeoNames codes them CAPE, STM and SHOR. The repeat control is the clearest single reading: WALLACHIA, which he had already called a normal region of Romania, he scores 85 where the scorer says 57.

Two consequences, one of them a correction to this page. The figure above that three of 41 new place names clear 70 is not a measurement of the class — it is a floor set by a scorer now known to sit about 38 points low. It is left in place rather than rescaled, because sliding every score up by the mean error would be fitting to nine words and would still not fix the ranking. And the anchor set has a hard limit worth naming: all nine are accepts, so it can calibrate the level but cannot locate the cut. Words he would refuse are what the next ask needs.

The deeper answer is that the missing factor may not be mechanisable at all, which was his point. The dump carries no area, so a magnitude signal has to come from somewhere, and the obvious candidate — the length of a place's alternate-names list, on the theory that the world names major things in many languages — was measured rather than assumed. It half works: Alaska Peninsula carries 30 alternate names against McGinnis Lake's 2 and MacInnis Lake's 0, which is exactly the separation wanted. But ADÉLIE COAST carries none at all and Eric scores it 80, and Lake Superior carries 10 against Lomita's 21, so the proxy inverts on the very cases it was invented for. No field in a 13.47-million-row gazetteer distinguishes a continental peninsula from a creek. That is why his instruction was a reasonableness standard rather than a tier, and the honest reading is that this one sub-problem belongs to judgement with the gazetteer confirming existence — the opposite of the direction this project has been pushing, and his call to make.

Three more rulings arrived overnight, and each changed something concrete. On alternate names: they generally should be fine … maybe just downscore an alternate if it has extremely low frequency corpus attestation — so the discount is now conditional on the corpus rather than flat. That is a better rule than either version this page has carried: a flat penalty docked ARGENTINA for arriving as an alternate of Argentine Republic while docking a genuinely obscure exonym no more, and his test separates those two cases using a fact already on disk. Applied, it lifts RAJASTHAN, KARNATAKA and CHONGQING to his line, since each reaches a state or municipality of tens of millions through an alternate carrying heavy corpus attestation. That reading needed one correction, found by the crossword project running the same scorer over three-letter candidates. Reading the corpus count for a string as attestation of the name only holds when the string is long enough that its occurrences plausibly refer to the place: ARGENTINA's 39 million rows really are about Argentina, but hnd has 674,090 — Higher National Diplomas and Haneda airport — and not one of them is evidence that HND names India. The count waived the discount on a three-letter code and carried it to India's own country row at over 70. So the waiver is now length-gated, and a lowercase string of four letters or fewer with no vowel is treated as a code outright, which is what HND is. The cost is stated rather than hidden: a genuine short exonym now takes the discount too, so RIO scores 47 where it deserves more. Below six letters corpus frequency cannot tell an attested exonym from an airport code, and the conservative error is the right one. The fix is inert on this project — every candidate here is nine letters, so the gate never fires — and the 116 scores are unchanged. He also pushed back on the question itself — wallachia is a normal region of Romania — what's the issue? — and he is right: the oddity was GeoNames filing it under Muntenia, not anything doubtful about the word.

On the archaic class, the narrow reading is confirmed: just unattested. All 177 refusals stand, because every one is of the publisshe type the judging model produced rather than a historical form with a citation. One thing worth recording about that ruling: the four examples he refused all carry corpus probability at or above 0.7, so corpus attestation cannot be used as the test of attested here — his own refusals would pass it. Distinguishing a cited archaic spelling from an invented one needs a dictionary this pipeline does not hold, and that limit is stated rather than papered over with a frequency proxy.

And he asked for the compound sample: send the 20. Building it caught a gap of our own. The 913 unruled compounds included 17 strings whose own clue calls them a misspelling — stratedgy, Misspelling of "strategy", had reached the draft sample — so they are now their own class and refused without asking anyone, which is the right use of an expert's time and takes the rule-refused total to 563. The 20 that remain are spread evenly by rank across the 0.80–0.93 band where his own six examples contradict each other, and one of them is THIGHBOOT, which he has already ruled in: it landed there by the spacing rather than by design, and it is useful left in as a check that his line holds between sittings.

He scored all twenty, and the result is that the line this page promised to fit does not exist. He accepts 16 of 20, refusing only villamare (65), narroways (65), madarasas (60) and bootleged (40). The offer made to him was to apply whatever line your 20 verdicts imply to the rest — so the first thing to do with the verdicts is test whether such a line is there, before a single one of the other 877 is admitted on it. It is not. The rank correlation between corpus probability and his score across the twenty is −0.09, and no threshold beats the base rate: the best-scoring cut is admit everything, which reaches 80% only because he accepts 80%. Every real threshold does worse — 0.81 scores 50%, 0.86 scores 45%, 0.89 scores 30%. A line fitted from these twenty would be a guess wearing the clothes of a measurement, so none is being fitted.

That is the third independent confirmation, on a third dataset, that english_probability does not discriminate validity within a band: +0.05 across his first scored list, a perfect inversion across his six compound examples, and now −0.09 across twenty scored compounds. The field measures how much a string resembles English text, and that is simply not the question.

A fourth dataset narrows that claim, and the narrower version is the true one. Eric's 0–5 grades for Grant's sixteen give a rank correlation against english_probability of +0.45 — positive, and on sixteen points not separable from chance at the conventional two-tailed level, where the 5% critical value is about 0.50. It stood at +0.53 until THAMNATA's regrade moved a single point; one word shifting the coefficient by .08 is its own warning about how much sixteen observations can carry. That does not rescue the field; it locates where the field is useless. Those sixteen span the whole range of wordhood, from a participle with 58.9 million occurrences to a biblical toponym with 149, whereas all three earlier datasets were restricted bands by construction — the twenty compounds sat inside 0.80–0.93, the six examples were all well-formed compounds — and restricting range destroys correlation. So the honest statement is that ep carries moderate signal across the full range of English and none inside the narrow band where candidates actually sit, and only the second half decides anything, because a gate is only ever asked about words near its own threshold. An earlier version of this page said flatly that the field does not track validity; that was too strong, and this paragraph is the correction.

The gate test is untouched by it, and still the more damning number. Of the three entries Eric grades 4 or 5, the 0.70 floor drops two — trattled at 0.68 and tropisms at 0.58 — and keeps one.

What the twenty do license is a rate rather than a rule. 80% of the 0.80–0.93 band clears his cut, so of the 228 compounds in that band roughly 180 would be valid entries by his standard — though on 20 words that proportion carries an interval of roughly 56–94%, which is 130 to 215 words, and the honest version of the claim is the interval rather than the point. The 663 compounds below 0.80 were not sampled at all, and nothing here speaks for them; sampling that end is the next cheap thing to ask him for, and it matters more than the band, because it is three times the size.

Two checks on the process itself came out of the same twenty. The control worked: THIGHBOOT, which he had ruled in previously, came back at 80. And one word he has now ruled on twice disagrees with itself at the boundary — hardhouse, which he eyeballed in his first reply as looks like OUT, scored 70 here, which is exactly his accept line. His glance and his scored judgement differ at the margin, which is the reason the band had to be scored rather than estimated from his examples.

A separate test, of the judge rather than of the words. Another model proposed ten "morphological gaps" — words it argued should exist but are absent from dictionaries. Eight were the wrong length for a 9×9 and so could not enter the search whatever their merit, but they made a clean probe of whether this project's judging prompt catches plausible invented morphology, since the prompt was written against programming identifiers and scanning noise. Run through the oracle, the panel accepted six, marked two marginal and rejected two. Eric then ruled on the same ten: 1,3,4,5,6,7,9 understandable / 2,8,10 bogus — bogus being TERRID, COMBUSTLESS and SINGERING. The judge and the expert agree on nine of ten, with a single over-admission: COMBUSTLESS, which has no corpus row anywhere in 21 million and was marked marginal rather than refused. The prompt also caught SINGERING for the right reason, writing its own clue as (Circular: to singe something) — the circular-clue rule firing as designed. So the judge is in better shape on this class than expected; the gap it has is an attestation rule, and the pipeline already holds the corpus count needed to enforce one.

Where the square actually dies

Everything above concerns which words are allowed in. This section asks the prior question, which this project had never measured: can a 9×9 double square exist over this lexicon at all, and if not, at which row does it fail? The answer does not need a search. Fill the square row by row; after k across words are placed, each of the nine columns holds a k-letter prefix, and every one must begin some nine-letter word or that column can never be finished. So the question is a property of the word list, and it can be measured in minutes rather than waited out for days.

Measured by Knuth's estimator — walk randomly from the empty square to a dead end, multiplying the number of legal rows at each step, and average over many walks — the collapse is abrupt and it is early:

mean legal next rows at depth lexicon 1 2 3 4 5 curated 63,091 63,091 2,117 396 0.45 0.00 + judged 16,828 79,919 74,150 642 0.05 0.00

The wall is row 4. A typical three-row stack has fewer than one legal continuation, and no walk of 300 on either lexicon reached nine rows. And the comparison between the two rows of that table is the finding that matters: the judged tier widens rows 2 and 3 by about thirty-five times and makes row 4 nine times narrower. It buys breadth at shallow depth — which is exactly the ten-fold slowdown measured above — while the place the search actually dies gets worse, because the extra vocabulary opens a mass of exotic three-row stacks that have no fourth row at all. That is the a tier is a multiplier on search time, not free vocabulary lesson, finally quantified at the depth where it decides the outcome, and it predicts the running sweep returns empty.

A second search went looking for the rare fertile corner instead of the typical case, keeping at each depth the stacks with the most continuations rather than sweeping start words in order. It gets further — five rows — and then stops, and what it prints at the wall is the useful part, because stage 1 of a Borgmann search is supposed to hand stage 3 a pattern to fill:

sasararas a sixth row needs, by column: tetramera l · lr · o · ls · o · i · nr · a · r attacapan thibetans which admits exactly eight strings: bemegride lloloinar lloloirar llosoinar llosoirar lroloinar lroloirar lrosoinar lrosoirar

All five of those rows come from the curated list, not the judged tier, so this is not a provenance problem. And all eight candidate sixth rows are unpronounceable. That is a different and harder failure than running short of vocabulary: the generate step of a Borgmann search can only help where the pattern admits a possible word, and here it admits none. Four of the nine columns are already forced to a single letter, so there is nothing for a judge to rule on and nothing for a corpus scan to find — a scan of all 21 million rows against the equivalent pattern one row shallower returned zero of its 1,980 combinations.

This is evidence rather than proof: the beam explored a thin slice, and a wider one might reach a fertile five-row stack whose sixth row is at least sayable. The exhaustive sweep is the proof, and it has fifteen days to run. But the measurement says what it will find, and it relocates the problem. The missing ingredient is not words a logologist could rule on. It is nine-letter words whose letter structure interlocks, and the lexicon's own prefix statistics say those run out at row four.

The wider beam, and what the generate step actually bought

The paragraph above promised a wider beam, so here it is, with a better heuristic. Ranking partial stacks by their number of legal next rows is what walked into that gibberish: a stack can have many legal rows and still have starved a column down to one letter. The score is now the sum over columns of log2 of the letters each column still admits, so a stack that is merely good everywhere beats one that is spectacular in eight columns and dead in the ninth. The beam then ran eight times wider, sampling sixteen times as many children at depth 2 — 199 million of them — in 44 MB and under three minutes.

The verdict did not move. Row 4 is still the wall, and every five-row stack the wide beam reached is dead at row 6. What changed is the quality of the logjams, and that is what made the rest of the method testable: the best four-row stacks now leave a fifth row whose pattern admits three million letter combinations with no column forced, against eight gibberish strings before. A pattern that open with only one or two listed words fitting it is precisely the case a Borgmann search exists for.

So stage 2 ran against it, and the corpus did deliver real words the lists lack. Against one four-row stack the 21-million-row corpus offered psalliota (14,258 occurrences), the old genus of the field mushroom, now Agaricus; others produced dormeuses (523), bornouses (365), dormouses (137) and caracores. These are the middle tier the method is named for — attested, clueable, absent from both word lists — and they are worth adding whatever happens to the square, though not while a sweep is running on a frozen lexicon.

Each one was then placed and the next row recomputed, which is the step that decides it. Every candidate survived its own row and killed the one after. Placing PSALLIOTA under CREMASTER / HEMOCOELE / ALBORADAS / TERRORISE leaves a sixth row that must read h s · a · n · · · — four of nine columns forced — and no English word begins hs. The corpus agrees: zero of those 324 combinations occur in it.

One column is worth looking at, though not for the reason this page first gave. With PSALLIOTA placed, column 1 reads chatp, and exactly one nine-letter word in either list can finish it: chatphoto, from the judged tier — labelled real, clued as "an image shared during an online conversation", with 1,669 corpus occurrences. A down entry is not disqualified for being unlisted, or even for being coined. The columns of a double square are entries on the same footing as its rows, and Borgmann's method admits clueable coinages in either direction; Grant's 8×8 was solved that way, which is why this project uses the technique at all. What matters is how many non-standard entries the finished square would hold and whether each clears the expert's line — and a de-spaced internet compound is exactly the class Eric scores unevenly, four of his twenty falling below his cut. So CHATPHOTO is a question for him, not a refusal. It is also a clean demonstration that the 0.7 corpus floor is no judge of quality: it ranks CHATPHOTO at 0.75 above PSALLIOTA at 0.50, a genus name with 14,258 occurrences.

What kills this stack is the row, not the column. The sixth row admits 324 strings, none of them in the corpus and not one that could be clued as anything in any language — a coinage still has to be sayable and definable, and no English word begins hs. Which is the honest result of running the method to its end: the generate step works, and it found real words the lists lack, but it can only help where the pattern admits a possible word. The deficit is not vocabulary a logologist could rule on. It is interlocking letter structure, and no amount of judging supplies it.

Trying harder, and finding out what a negative result is worth

Two further searches were built, each fixing a specific defect in the one above.

sqdeep removed the random sampling and changed what "healthy" means. The beam had been keeping 60,000 of depth 2's 200 million children at random, so a promising stack had a three in ten thousand chance of being looked at; children are now scored and kept per parent, so nothing good is lost to a coin toss. And column health is now counted in words rather than letters — a column admitting one letter with five hundred words behind it is healthier than one admitting four letters with four words behind them — which makes the score the log of the product of per-column options. That took the search from four rows to six, and depth-first search then drove all 92 six-row stacks to the bottom with no beam at all: no square below any of them. The corpus had nothing for any seventh row, and the patterns admit nothing clueable.

sqalt changed the order of play. Every search here, the exhaustive sweep included, fills row by row and tests a column only by whether its prefix still begins some nine-letter word. That test is necessary but weak: it passes a stack where each column could be finished separately while no choice of rows 7, 8 and 9 finishes them all at once — which is exactly how those 92 stacks died. So sqalt places row 0, column 0, row 1, column 1, and onward, committing each down word as a real word the moment its letters are known. It reached twelve of seventeen lines — six across words and six down words at once — in the first twenty-five seconds, and then spent 728 million further nodes without ever placing a seventh across word. 738 million nodes, 410,000 a second, no square.

Which raises the question this project had not asked: what is a negative result from a heuristic search actually worth here? It can be measured. sqseed fills a 9×9 grid with random letters, reads off its nine rows and nine columns, and writes those eighteen strings as a word list — a list that contains a double square by construction. Pad it with real words and the solver has a needle it is guaranteed to be able to find:

lexicon size planted square found? 18 yes, in 18 nodes 218 yes, under a minute 2,018 yes, under a minute 8,018 yes, under two minutes 20,018 NO -- stalls at 8 of 17 lines 40,018 NO -- stalls at 10 of 17 lines 79,919 (the real lexicon)

The ability to find a square that is certainly there collapses between eight and twenty thousand words, and the real lexicon holds 79,919. Inverting the heuristic to prefer thin prefixes, or discarding it for random order, does not help — all three orderings succeed at 2,000 words and all three fail at 40,000. So sqalt finding nothing in 738 million nodes is weak evidence about existence, and the six-row barrier is where these searches stall rather than where squares stop existing. Saying otherwise would be reading a limit of the instrument as a fact about the world.

What survives is worth separating carefully. The evidence that no square exists over this lexicon is the Knuth estimate — unbiased, sampling-based, and far below one on both word lists. The evidence that a search would not find one anyway is the table above. And the only thing that will prove it either way is the exhaustive sweep now running, which is why it is still running. The planted-square benchmark also gives the next solver a target that is not a matter of taste: find the square at 79,919 words.

Its cost was measured rather than assumed, and measured the only way that is sound here — both lists at once, on the same eight windows of the start-word range, on the same box, because the cost of a start word varies sixteenfold across that range and a single window proves nothing.

window 9,454 words 16,828 words ratio 0 7 2 0.29 8,000 35 3 0.09 16,000 51 4 0.08 24,000 39 5 0.13 32,000 26 2 0.08 40,000 28 1 0.04 48,000 14 1 0.07 56,000 60 5 0.08 TOTAL 260 23 0.09 start words/hour, 18 lanes: 3,510 vs 310

Two things in that table are worth more than the ratio. The first is a check on the whole method: the trimmed list's measured 3,510 start words an hour matches the 3,457 the real eleven-hour sweep actually averaged, so this bench predicts reality to within two per cent. The second is mechanistic. Both halves of the benchmark explored the same number of nodes — 181 trillion against 172 trillion — in the same ten minutes. The richer lexicon is not slower per step; it makes eleven times less progress per step, because every added word lets more partial squares survive pruning at every cell, so each start word's tree is eleven times bigger. Vocabulary does not cost overhead. It costs search space.

So the full-tier sweep is eight and a half days rather than the eleven and a half hours its trimmed sibling took, and it started on 16 October. If it also comes back empty, it supersedes the result above rather than repeating it: no 9×9 asymmetric square with a curated word in its top row or left column exists even when every judged word is allowed in everywhere else, coinages included.

Correction, 25 October: what the sweep covers. Earlier versions of this page said the full-tier null would hold for every square over the judged tier. It does not, quite. The solver places each start word as the top row and fills the other eight rows from the whole lexicon, but start words come only from the 63,091 curated words; the judged words are fills. By transposition a square is therefore found when its top row or its left column is curated, and a square whose top row and left column are both judged words is never visited. A planted-square test with the sweep's own binary and exact word lists confirms both halves of that: a square planted with its top row curated is found in 28 seconds, one with only its left column curated is found transposed in 61 seconds, and a negative control finds nothing. The remaining squares get a second, much smaller pass — judged words as start words, with the left column restricted to judged words too, so the two passes meet without overlap. That pass has been proved on the same planted square before running anything real: with all eighteen planted words as judged words it finds the square from its top row in 28 seconds and, transposed, from its left column in 37; with one row withheld it finds nothing; and for a square whose left column is curated — the main sweep's territory — it correctly finds nothing too. Until it has run, the claim is the narrower one above. If instead it finds one, the square has to be read off its provenance block before anyone celebrates — every row is marked base, fill or oracle, and a square resting on one of the 431 coinages is a question for a human expert rather than a record.

The pass the 0.7 corpus floor had made unaffordable ran anyway, on nuc13 over the full 16,828-word judged tier with no corpus filter at all. Its 400 chunks took 17.4 days, averaging 151 start words per hour against 3,457 for the trimmed list on the same machine — the cost of vocabulary, not of overhead: both lists explore nodes at the same speed, the richer one just leaves more partial squares alive at every cell.

What the lanes have done so far implies nothing left to search remaining — complete — with 100.0% of the range swept. These are the live figures on this page: the clean pass above is finished and its numbers will not move again.

The Borgmann search: growing the lexicon where the search dies

The sweeps above prove nulls over a fixed lexicon, and the full-tier one is now kept for completeness. If a 9×9 double square exists, it most likely needs words nobody has listed yet, proposed and judged at the exact places the search runs out — a true Borgmann search. Since 25 October that is the main effort. The search places a start word as row 1 and fills further rows one at a time; a dead end is a partial square where no word in the lexicon fits the next row, and its pattern is the letters each cell of that row could still take (for example [an]tten[ftz]is[rt]). Any word that fits the pattern lets the search go on; at row 9 it completes a square.

The first measurements set the direction. The search dies at row 6: over forty start words, all 10,829 partial squares that reached five rows were dead ends. Local data almost never has the word: only 9 of 10,674 row-6 patterns, and 878 of 130,184 row-5 patterns, fit any unlisted string among the 2.56 million nine-letter forms in the 21-million-entry corpus, and most of those are typos. Even admitting every corpus string that occurs five times or more, unjudged, carried no partial past row 7 in a 30-billion-node sample. That measures the corpus, not English — and the unreliability of corpora is Borgmann's whole point, in both directions: the corpus is full of strings no one would accept, and it lacks most of what a logologist would. So the loop does not ask the corpus. At each dead end it lists, in C++, every string that fits the pattern, and has each one judged on Eric Chaikin's scale — as a word, a phrase, a name, an initial and a word, a run-together form — with nothing filtered out beforehand. At the first 101 six-row dead ends that was 881 strings, and the judge admitted none of them.

The Borgmann search is NOT running on nuc13 (started 190.8 hours ago), on four lanes that the load governor serves before the sweep: 400 of 400 chunks scanned (about 63,091 of the 63,091 start words, each searched to a budget of a billion nodes), every dead end from row 6 down kept with the exact letters each cell of the next row could take. No square yet.

Also complete, on titanv's TITAN V GPU: word_square_gpu over the older 116,487-word Borgmann list of 10 July (its t80 list plus phrases), a separate search from the nuc13 passes above. It has searched all 116,487 start words, and the program reported 0 verified solutions. That list predates Eric Chaikin's grading, so the result is for the list as it stands, junk entries included.

Two things make the null result worth more than its own figures. The vocabulary it ran on was one enumerable file, not a service that could answer differently on a second run, and the md5 of that file was re-checked before every lane started. And the one check that could have invalidated it was carried out rather than assumed: phase 2's 15,905 additions were sorted against the frozen list after the fact, and nothing solid was found to be missing. What would move this result is not more compute — it is vocabulary a human expert will vouch for.

The two words the sweep cannot reach, answered without touching it

While the full-tier pass runs, its vocabulary is frozen by an md5 check at every lane start — and that is not fastidiousness. A chunk checkpoint only means something in the search space it was written in, so a lane that comes back after a reboot must come back to the same lexicon. Worse, a third of the 400 chunks are already finished; a word added now would be absent from everything they covered, and the pass would end as a null over neither the old lexicon nor the new one. So a word earned on 19 October has to wait until about 28 October — or be answered another way.

It can be answered another way, and much more cheaply. Every square the sweep can find lies in the old lexicon. Every square it cannot find must contain one of the new words. Fix such a word into a row, fill the remaining eight rows exhaustively, and the sweep plus those pinned jobs cover the enriched lexicon with no overlap and no gap. Two details carry it. Rows alone suffice: a square whose column is the new word is the transpose of one whose row is, and the transpose of a double square is a double square over the same list — nine jobs per word, not eighteen. And the pin prunes upward, which is what makes it fast: with row p fixed, column j must be spelled by a word whose p-th letter is already known, so the mask consulted at every cell is not "letters that continue this column" but "letters that continue this column towards such a word". Without that, a pin at row 7 would prune nothing until the search reached row 7.

The first thing the exercise produced was a correction. Five words had been earned from the corpus scans; only two of them were new. words_9letter_cat1.txt holds 86,692 lines and the solver loads 63,091 words, because its loader lowercases, strips non-letters and discards duplicates — so PSALLIOTA (present twice, in both cases), caracores and dormeuses were in the search space all along, and the running sweep covers them. A grep on the file and a question about the search space are not the same question.

That left bornouses and dormouses, and 18 jobs. First, though, the instrument had to be shown to work — and the planted-square benchmark above sets the bar: find a square that is certainly there, at the real lexicon's size, which no heuristic search here has managed past 8,018 words.

planted square, delivered through the delta file result 18-word list, pinned at its own row found, 83 nodes wrong row pinned (negative control) exhausted, no square, 3 nodes 20,017-word lexicon found, 1.15 million nodes, under a second 79,937 words: planted 18 + cat1 + full tier FOUND, 62 s, 2.25 billion nodes

That is the target met. One known line is worth more than every candidate ordering tried in the section above — thick-prefix, thin-prefix and random all failed at 40,018 words, and the pin succeeds at twice that. Then the real jobs, against the 79,921-word enriched lexicon:

pinned row bornouses dormouses 1 4.86 bn nodes 249 s 6.13 bn 295 s 2 9.66 bn 451 s 4.86 bn 216 s 3 8.09 bn 363 s 8.37 bn 397 s 4 18.97 bn 866 s 18.93 bn 906 s 5 51.36 bn 2034 s 50.32 bn 1865 s 6 66.52 bn 2845 s 56.21 bn 2733 s 7 86.99 bn 2056 s 71.68 bn 2843 s 8 26.74 bn 888 s 74.63 bn 2235 s 9 0.80 bn 25 s 7.85 bn 273 s every job: exhausted, no square

Delta search · complete

Neither bornouses nor dormouses can appear in a 9×9 asymmetric double square over the enriched lexicon — in any row, and therefore in any column. All eighteen jobs exhausted their subtree; none was cut short.

573 billion nodes · 5.98 core-hours · 0 solutions · 20 Oct, on the desktop, nuc13 untouched

The cost is the point. The sweep has spent close to 200 trillion nodes so far and needs eight more days; this answered the same question for these words in three thousandths of that, on four cores that were doing nothing, while the sweep ran undisturbed on the other machine. Re-sweeping to bank a word is never the right move again.

The shape of the cost is worth recording too. Pinning near either end is cheap — row 9 for bornouses took 25 seconds, because the final row's letters constrain all nine columns at once — and the middle is dear, peaking at rows 6 and 7 where the pin is too deep to prune the early rows and too shallow to finish them. That peak sits exactly where the beam searches stall, which is the row-4-to-row-6 wall seen from the other side.

What this does not say: anything about squares that use neither word. That remains the sweep's job, and the sweep is still running.