Dated reports from Chris's search for symmetric 15×15 crossword grids with a five-deep stack of grid-spanning entries, written by Claude for Eric Chaikin. Newest first.
Rows from the web: one weak fill, and scoring words by use
Chris's Windows project has been counting every word written on the web, from Common Crawl, a public archive of billions of web pages. Its count finds words that no dictionary lists, which Chris expects to be the likeliest source of good stacks. Today the first of them went into the search as stack rows. The result is one weak fill and no grid to send. It also raised a question about how every entry should be scored.
881web words tried as the first row of a stack
285layouts found, on 13 stacks
1stack that fills at 50, with eight entries below 50
1,373more web rows waiting for usage scores
The first web rows
The first count covered 2,000 of Common Crawl's 100,010 files. It had 1,286 fifteen-letter words that are on at least 10 sites and that our vocabulary lacks. 947 passed the same LLM screen as the earlier rounds. I read every pass and dropped 63 that are not entries: code names, foreign words, spam, game items and misspellings. That left 881, almost all single technical words such as DISCOVERABILITY, BACKPROPAGATION and VERMICOMPOSTING. Phrases are counted only where sites write them solid, so no phrase like PUTS THE KETTLE ON came from this count.
Searched as first rows on the GPU, the 881 gave 285 layouts on 13 stacks. 282 layouts were rejected before any search, because some slot had no possible word. No layout filled with 0 to 6 entries below 50. With 8 allowed, one stack filled, at Q 1.806:
Stack row
Score
RUMINOCOCCACEAE, a family of gut bacteria (web row)
unjudged
ACCOUNT SETTINGS
75
TOILET ATTENDANT
50
ANTECONSONANTIC
40
SERPENTINE VERSE
40
Three of the five rows are below 50. So are five crossings: CTENES, EUCONE, NUECES, CONTACTANT and SODIUM CITRATES. It does not come close to grid 12 (Q 1.822 with nothing below 50) or grids 9 to 11, so I have not sent it. The other twelve stacks, on rows such as NANOFABRICATION, PHOTOBIOREACTOR and SOMATOSENSATION over ONE MOMENT PLEASE or AS LOOSE AS A GOOSE, did not fill at all.
The interim web list
This evening the count covered all 100,010 Common Crawl files. It has 38,633 fifteen-letter words that are on at least 10 sites and that our vocabulary lacks. Most are hyphenated adjectives such as STUDENT-CENTERED, plus foreign words, casino spam and names. New words written solid are rare at this length: the most used is on 567 sites. Two-word strings of 15 letters are ordinary word pairs such as UNDERSTANDING OF, so I did not count them as candidates.
The 3,063 words on at least 100 sites went through the LLM screen, and 1,822 passed. I read them all and removed 439 that are not entries, and 10 had already been rejected by our judge. That leaves 1,373, among them STUDENT ATHLETES, CROSS VALIDATION, CIRCULAR ECONOMY, OXIDATIVE STRESS, VICTIM MENTALITY and DEFUND THE POLICE. Most, though, are words only a specialist would know, such as HEMILAMINECTOMY. They are in the vocabulary file at a provisional score and have not been searched.
Scoring words by use
Chris pointed to Dmitri Borgmann's complaint that dictionaries leave out words like AIRVENT, which every English speaker knows, because the meaning is obvious. The web count bears it out: air vent is on 16,130 sites, air vents on 39,096 and airvent, written solid, on 304, yet none of the dictionaries in the count lists it. Our crossword lists give it 50, the bare pass mark. Chris's view is that leaving such words out was never justified, by saving space or anything else. The aim instead is a way of recognizing words the way English speakers do, tested against the words they actually use on the web.
That matters here, because the scores decide which grids count as finished. The newer tiers each sit at one flat score, and ANTECONSONANTIC scores 40 because it is in a word list, not because anyone uses it. The plan:
Build a test from the web count. The positives are unlisted words in wide use; the negatives are misspellings, run-together strings and combinations no one uses.
Compare an LLM recognizer with one that builds words from known parts, and with the two combined.
Calibrate the best one against Eric's grades, on the same held-out half as before.
Score every entry by it: word-list entries, canonical-only entries, LLM rows and web rows.
Chris has moved this work to his wordlist project, which uses the Windows project's counts. Its first test has already run, on two-part words. There were 119,356 unlisted words in wide use, each paired with an unused combination of equally common parts. A model that builds words from known parts told them apart with an AUC of 0.70, where 0.5 is chance and 1.0 is perfect. Almost all of its signal came from how often each part forms other words. The LLM recognizer has not yet been scored on the same test. When its scores are ready, I will rescore the vocabulary and recheck the layouts already found. The 1,373 new web rows wait for those scores before they are searched.
What is running
The full rotational search with the round-3 and round-4 rows, described below, paused twice today for other GPU work. It is running again from first row 2,894 of about 45,500, so it will end somewhat after 18 November.
No new grid at 50; 385 more LLM rows go into rows 2–5
The rotational search with the 395 round-3 LLM rows has finished its last part: first rows 0 to 17,256, which had been searched before those rows were in the vocabulary. Nothing it found beats grid 12 (Q 1.822) or grid 10, the two rotational grids at the 50 standard. A new round of 385 LLM-proposed rows is now in the search, this time in every row of the stack.
194rotational stacks found on first rows 0 to 17,256
2of them use a round-3 LLM row; neither fills
14,367layouts checked at the 50 standard
0layouts that fill with no entry below 50
The check at 50
Most layouts, 11,431 of 14,367, were rejected before any search, because some slot had no possible word. Checks ran out of time at 60 seconds on one layout at 0 entries below 50, one at 1, two at 6 and seven at 8.
Entries allowed below 50
New fills
Stack and Q
0
0
1
2
ROTISSERIE GRILL / APOSTOLIC PALACE, grid 9's stack, best Q 1.811 with LESTS its only entry below 50
2–3
12
The same stack, best Q 1.809
4
0
5
136
Best: PUTS THE KETTLE ON / APOSTOLIC PALACE / SENTIMENTALISES / TRIAL ENGAGEMENT / STEREOKINESISES, Q 1.823, the same fill as in the 6 November entry
6
182
Best Q 1.814
8
1,043
Best Q 1.816
So this last part found the same best stacks as the earlier search of these first rows. The two stacks with a round-3 row, on THOMSENS DISEASE and ENTERPRISE GRADE, did not fill even with eight entries below 50.
Round 4 of the Borgmann search
The LLM was asked for real 15-letter phrases in 50 new areas, three of which failed twice. It returned 1,009 different 15-letter strings: 329 were already listed and 679 were new. 385 passed the screen. They are candidates only: none has been web-checked or graded, and they enter the vocabulary at 42, marked so that any grid using one shows it.
Searched as first rows, the 385 gave 122 layouts, and all 122 were rejected before any search because some slot had no possible word. The same round asked for entries in the slots that no word fits. For the long downs through the stack, splitting the pattern in two and asking for each half gave one real entry, SEA ORGANIZATION, and it fills none of the 36 layouts it fits. For the short corner slots the proposals were only abbreviations; I have held back all eight rather than spend Google checks on them.
Adding the 15,851 words new in the NWL2023 and CSW24 Scrabble lists at 40 changed no fill at 0 to 6 entries below 50. At 8 it added 16 layouts, all on one stack and all through the single new word IDPOL, at Q 1.762 to 1.782. The search's limits are not in the canonical lexicon.
What is running
Since this morning the GPU has been running the whole rotational search again, with both the round-3 and round-4 rows in the vocabulary, so that a round-4 row can be row 2, 3, 4 or 5 of a stack. That is about 45,500 first rows, roughly 100 GPU hours at the last part's pace, so it should end around 18 November. Each first row still tries only its 1,000 best second rows, so it is not exhaustive.
Grid 12, Q 1.822: a rotational grid at the 50 standard with nothing to grade
Every entry in grid 12 scores 50 or more on the list scores and Eric's grades, so it meets the 50 standard with no word to grade. It is the second rotational grid at the standard, after grid 10 (Q 1.810), and its Q of 1.822 is the higher. Every entry comes from the word lists: none is LLM-proposed and none is canon-only.
312rotational stacks found on first rows 26,003 to 45,506
18,833layouts checked at the 50 standard
6layouts that fill with no entry below 50, all from one stack
0words for Eric to grade in grid 12
Grid 12 · Q 1.822 · rotational symmetry · at the 50 standard
Stack rows: VIBRATORYROLLER, ONLINEPETITIONS, LEASTRESISTANCE, HASTHELIMELIGHT, ALTERNATENESSES. Columns 3 and 13 are full-height downs.
Entry
Score and note
ALTERNATENESSESstack
70. The weakest stack row: a plural of an abstract noun.
VIBRATORY ROLLERstack
56. The road-building machine.
MYELOBLASTEMIAS
66. A full-height down through the stack; a medical plural.
JAMES LONGSTREET
75. A full-height down: the Confederate general.
LVER
50. On the list, but I know no meaning for it. It sits in the top-left corner, which can be reworked.
DMME, IIIS, ESKS, ABLIP
75, 50, 50, 60. Weak short entries in the corners.
Fourteen entries sit at exactly 50: ADJS, DEYS, DOTTLE, ESKS, EVARISTE, IIIS, LIAISING, LVER, NANO, RERISEN, RESIT, SEDE, SUNDRENCHED and TERENCE. The corners lean on weak short entries. Chris tells me Eric has always been able to replace those with better ones in the same area, so what counts here is the stack and the long downs locked to it. The clues in the PDF and .puz are my own first drafts.
Grid 12 comes from the last third of the search, which also found the same stack in other layouts. It fills with one entry below 50 in 11 layouts, LIBRISTE or ANTHRA (both canon-only), and with none in 6 of them. Grid 12 is the best of those six.
The check at 50
Two checks covered every layout found on first rows 26,003 to 45,506. Four layout checks ran out of time at 60 seconds; they will be rerun with more time.
Entries allowed below 50
New fills
1 (6 of them also fill with 0)
11
2
25
3
19
4
134
5
286
6
316
8
424
The fill check stops at the first fill it finds and began at one entry below 50, so the six fills with none showed up only when a new step searched each layout for its best fill. That step comes from Orca, an open-source grid filler, which lists every fill of a grid rather than stopping at the first. The new step keeps searching after a fill and keeps the best one. A second Orca idea, remembering rows that fitted before, made the GPU search 8% slower and is not used.
Limits
The search with the 395 LLM-proposed rows has now covered first rows 16,148 to 45,506. Each first row still tries only its 1,000 best second rows, so it is not exhaustive. Of the 395 LLM rows, only TRANSPORTER ROOM has been checked on Google, and none has been graded.
Grid 12 is also in grid12.pdf (grid, clues and answers) and grid12.puz, in the “Claude Artifacts / crossword-stack” folder on Chris's Google Drive.
Grid 11, Q 1.817: an LLM-proposed row, four words from the 50 standard
The 395 LLM-proposed rows from the last entry are now in the main rotational search, where they can sit in rows 2 to 5 of a stack. One of them, TRANSPORTER ROOM, is the middle row of grid 11. At Q 1.817 it has the highest Q of any rotational grid in this log, but four entries score below 50, so it is not at the standard. They are the LLM row and three words found only on the canonical word list. Grid 10 is still the only rotational grid at the standard.
84rotational stacks found on first rows 16,148 to 25,607
4,733layouts checked at the 50 standard
34filled with 4 or 5 entries below 50; none with 3 or fewer
4ungraded entries between grid 11 and the 50 standard
The room on a Star Trek starship where crew are beamed. Proposed by the LLM in round 3 and not on any of our lists. An exact-phrase Google search shows it on 6 sites, including Memory Alpha and a Star Trek podcast. That shows the phrase is in use, mainly among fans, not that it is a fair entry.
ANTIDETONANT
An antiknock additive for fuel. Canonical word list only. A 12-letter down.
MONOCEROSES
Plural of monoceros, an old word for a unicorn or other one-horned creature. Canonical word list only. Weak.
COPAN
Copán, the Maya ruins in western Honduras. Canonical word list only.
All other entries score 50 or more. Seventeen sit at exactly 50, among them three of the stack rows, BEDAMN, BRISTLE BRUSH, DOR, DSO, ESAR, ONORATI and TITULAR ROLE. DOR and ESAR look thin even at that score. The clues in the PDF and .puz are my own first drafts.
Two other stacks also fill with four entries below 50. Neither has an LLM row, and their weak entries are worse: SEEK A SENATE SEAT / ALASKA PANHANDLE / PECTINESTERASES / ON CONDITION THAT / REORGANISATIONS reaches Q 1.811, but it needs DEPEINCTING, FAINEANTISE, OOT and PECTINESTERASES.
The check at 50
Three checks covered every layout the search found on first rows 16,148 to 25,607. Most layouts were rejected before any search, because some slot had no possible word. Nothing timed out.
Entries allowed below 50
New fills
1 to 3
0
4
32
5
2
6
16
8
766
A faster search
The GPU was much less busy than its utilization figure claimed: it drew about 45 W of its 250 W while a few long jobs held each batch. A change that cuts a long job into many pieces for idle parts of the GPU to share made the search about 3 times as fast, with identical results on a 60-row test. Rows 18,472 to 25,607 then ran at about 470 first rows an hour, against about 150 before.
Limits and what is running
The search now covers first rows 26,003 to 45,506 and should take about two more days. Each first row still tries only its 1,000 best second rows, so it is not exhaustive. Of the 395 LLM rows, only TRANSPORTER ROOM has been checked on Google; none has been graded.
Grid 11 is also in grid11.pdf (grid, clues and answers) and grid11.puz, in the “Claude Artifacts / crossword-stack” folder on Chris's Google Drive.
No new grid at 50; 395 more LLM-proposed rows to test
Nothing today beats grid 10, which is still the only rotational grid from this search with no entry below 50. Grid 9 still misses by one word, LESTS. The checks below found no new grid at the standard, and a new round of LLM-proposed rows is ready to search. None of those rows has been searched, web-checked or graded.
0layouts from the 84 newly admitted rows
17new fills of 1,241 layouts, none at 50
395new row candidates passing the LLM screen, unjudged
13,510first rows reached of 45,111, 44 stacks found
The 84 rows admitted by a looser screen
Dropping the line “say NO unless you are sure” from the screen for long proposals let 84 of 408 earlier rejections through. Searched as first rows, they gave 152,954 stacks and no layouts. They were tried only as first rows, with second rows capped at 1,000, so this does not rule them out.
The line stays out. On 102 strings the judge had rejected, 10 passed without it and 9 with it, and 8 of those were the same strings. That sample is not pure junk, since several passes are real but obscure. Of 12 rejected strings checked by hand earlier, none passes either way, including two I rated clearly real, so the screen can still lose real entries.
The check at 50
1,241 more layouts: 17 filled with eight or fewer entries below 50, on 12 stacks. 1,194 proved inconsistent, 30 did not fill and none timed out.
ANTICORRELATION / LEONARDO DA VINCI / IMPORTANT PERSON / MARIE ANTOINETTE / STYLELESSNESSES, Q 1.776
6
0
8
15
Best: PUTS THE KETTLE ON / APOSTOLIC PALACE / SENTIMENTALISES / TRIAL ENGAGEMENT / STEREOKINESISES, Q 1.806
A correction to the previous entry: its check of 9,014 layouts missed 101 from the first part of the search. This check covered them, and they gave 7 of the 17 fills, including the Q 1.792 and Q 1.806 layouts above. The cause of the miss has not been established.
A new round of proposed rows
Round 3 asked the LLM for real phrases in 54 new areas, such as machinery, job titles, anatomy, regional foods and verb phrases. It returned 1,056 different 15-letter strings. 466 were already listed and 589 were new. Of the 565 not already screened in earlier rounds, 395 passed the screen. The passes include good candidates, such as HOLD THE WHIP HAND and LOSE TRACK OF TIME, and strings that are not real entries or would make poor ones, such as PILLOW CONCIERGE. They are candidates only. A row that reaches a good layout gets a web check, and then goes to Eric.
A way to search for a chosen row
A new version of the search can require a chosen word at a chosen row of the stack. On one known case, fixing the first row, each of the other four rows pinned at its own position found exactly the same 966 layout files as the unpinned search, in 8–51 CPU seconds against 44,139, and the wrong position found nothing. That is one case, not a proof. A sample of 20 LLM rows placed second in the stack has finished 25 of 200 blocks, with no layouts so far, which is too few to judge.
What is running
The main rotational search continues on the GPU, still taking only 1,000 second rows per first row, so it is not exhaustive. The current slice should end around midnight tonight if the pace holds, and the whole first-row list around 13 November. The 395 new rows are not yet in any search.
Eric's grades: grid 10 is the first rotational grid at the 50 standard, Q 1.810
Eric graded every word we asked about. All four words grid 10 needed came back at 3 or better, so on his grades and the list scores it has no entry below 50, the first rotational grid from this search to reach that. Grading the stack and the four words also moved its Q from 1.795 to 1.810. Grid 9 misses by one word, LESTS at 2.8. He gave grades only, with no verdict on either grid, and we have asked him whether he considers grid 10 finished.
0grid 10 entries below 50
1.810Q of grid 10 with his grades
3.3lowest of his grades in grid 10: VIBRATORY ROLLER
The other 57 entries have list scores of 50 or more, not grades from Eric, and 15 sit at exactly 50. LVER, IIIS and ESKS still look thin to me, and so does DM ME at 75. The 3.8 was written under a sentence naming three words; I read it as MYELOBLASTEMIAS, the one of the three he did not grade elsewhere. The clues are still my first drafts.
DESTROYING ANGEL 4 high, IN THE LINE OF DUTY 5 high, STARVING ARTISTS 5, ACTION COMMITTEE 4, SHIFTS ONES GEARS 3.8. LINEAMENTING “ok for now. Later we won't need these”; RENTCHARGE “ok”; UNTHRIFT, WSJ and RVS “all just ok”. No grade seen for GORMS.
The check at 50, with these grades
I then checked the 9,014 layouts the main search had written by this morning. 559 filled with eight or fewer entries below 50, on 13 stacks. 7,728 proved inconsistent, 727 did not fill and none timed out. None filled with every entry at 50, so grid 10, from the separate LLM-row pass, remains the only rotational grid at that standard.
Entries allowed below 50
New fills
1
2
2
3
3
8
4
0
5
135
6
142
8
269
The nearest is grid 9 itself (seed_0028_26). With his grades its Q rises from 1.794 to 1.811, and its only entry below 50 is LESTS at 2.8. LESTS is forced in this layout, because the last letters of the five rows spell it between two black squares. It has 52 black squares, against 40 in grid 10. All 13 fills in the first three rows of the table are on this stack. The highest Q is 1.823 on PUTS THE KETTLE ON / APOSTOLIC PALACE / SENTIMENTALISES / TRIAL ENGAGEMENT / STEREOKINESISES (seed_0006_54), with five entries below 50: CHIROTONIES, EOCENES, ILLIMITED, STEREOKINESISES and SST.
What is running
The main rotational search continues on the GPU, still taking only 1,000 second rows per first row, so it is not exhaustive. Grid 10 is the best rotational grid so far, not the end of the search. A sample is also testing whether LLM rows placed second in the stack, which that cap never reaches, can find further stacks.
Grid 10 is also in grid10g.pdf (grid, clues and answers) and grid10g.puz, in the “Claude Artifacts / crossword-stack” folder on Chris's Google Drive.
An LLM-proposed row: grid 10, four words from the 50 standard
A band search can only build stacks from 15-letter entries it already has, so until now the LLM could only repair crossings on stacks the word list allowed. This round it also proposed new 15s, and those rows were searched as first rows of rotational stacks. The best result, grid 10, needs four grades to reach the 50 standard, one fewer than grid 9, at about the same Q. It is not at the standard yet, and two of its four words are weak.
401LLM-proposed 15s searched as first rows, all unjudged
10rotational stacks found on them
137of 1,537 layouts filled with 8 or fewer entries below 50
4ungraded entries between grid 10 and the 50 standard
A road-building compactor. Proposed by the LLM and not on any of our lists. An exact-phrase Google search shows it on 6 sites, including Bobcat's. That shows the term is in use, mainly in the equipment trade, not that it is a fair entry.
ALTERNATENESSESstack
Plural of alternateness. Canonical word list only. Weak.
MYELOBLASTEMIAS
Plural of myeloblastemia, immature white-cell precursors in the blood. Canonical word list only. A 15-letter down, and weak.
VOLHARD
The Volhard titration, after the chemist Jacob Volhard. Canonical word list only.
All other entries score 50 or more. Sixteen sit at exactly 50: ACER, ADJS, BANE, DEYS, DOTTLE, ESKS, EVARISTE, HAS THE LIMELIGHT, IIIS, LIAISING, LVER, RERISEN, RESIT, SEDE, SUN-DRENCHED and TERENCE. LVER, IIIS and ESKS look thin even at that score, and so does DM ME at 75. The clues in the PDF and .puz are my own first drafts.
The highest Q on this stack is 1.8065 (seed_0005_28), but it needs six grades: ALTERNATENESSES, LIBRISTE, MYELOBLASTEMIAS, PINEALOMAS, VOLHARD and VIBRATORY ROLLER.
How the rows were made
Gemini 3.8 Flash was asked, one theme at a time over 100 themes, for real phrases of about this size, with no pressure on exact length, because asking for exactly 15 letters produced cut-off and padded junk. C++ kept those of exactly 15 letters that were not already listed. A frequency screen and an LLM check followed, then I removed 22 by hand, leaving 401. Google has checked only VIBRATORY ROLLER so far; the others are checked only if they reach a layout worth pursuing. So these rows are unjudged, and this entry's numbers say nothing about whether they are good.
The check at 50
The 401 rows gave 10 stacks. Every fill below comes from one of them, the VIBRATORY ROLLER stack. The other 1,359 layouts proved inconsistent, 39 did not fill, and two timed out at 60 seconds (seed_0005_234 and _235), so those two are unresolved. An earlier batch gave one further fill on an LLM row, GEORGE HURSTWOOD, needing up to eight grades at Q 1.783.
Entries allowed below 50
New fills
1 to 3
0
4
6
5
5
6
25
8
101
Limits and what is running
The main rotational search continues, and each first row tries only its 1,000 best second rows, so it is not exhaustive. Rows already searched before the LLM rows existed never saw them in positions 2 to 5, so that needs a second pass. The GPU machine hung on 5 November and the search lost about 40 minutes; it has resumed, with no results lost, and I expect it to take another 9 to 10 days.
Grid 10 is also in grid10.pdf (grid, clues and answers) and grid10.puz, in the “Claude Artifacts / crossword-stack” folder on Chris's Google Drive.
Rotational search: a candidate grid five words from the 50 standard, and two faults fixed
Following Eric's ranking of the symmetries, the band search now covers rotational symmetry only, with much wider sampling of each stack's second row. About 20% of the first rows are done. No rotational grid has reached the 50 standard yet, but the nearest, grid 9 below, needs only five ungraded canonical words. These words come from the larger lexicon, not from a Borgmann step.
108rotational stacks found so far
8,969of 44,710 first rows searched
0fills with every entry at 50, on the first 70 stacks
5ungraded words between grid 9 and the 50 standard
Students of church doctrine or, in an older sense, church architecture. Ecclesiology is well attested; the agent noun is a regular formation.
EPAGOMENAL
Of the five days added to the ancient Egyptian year, the standard term for them.
GALEGINE
The alkaloid of goat's rue, the starting point for metformin.
TAENIASES
Plural of taeniasis, tapeworm infection.
LESTS
Weakest of the five. Webster's 1913 has an obsolete verb lest, “to listen” (Chaucer, Spenser), but I found no citation of this form.
All other entries score 50 or more; seventeen sit at exactly 50, among them MYEAR, HED, UAP, LLENO, ERNESTOS and RASPE. An editor may also object to the runs of seven black squares along the top and bottom rows.
The nearest grid with a stronger stack is at Q 1.800: DESTROYING ANGEL / IN THE LINE OF DUTY / STARVING ARTISTS / ACTION COMMITTEE / SHIFTS ONES GEARS. It needs six words below 50: GORMS, LINEAMENTING, RENTCHARGE, UNTHRIFT, WSJ and RVS. I found LINEAMENT only as a noun, so LINEAMENTING is probably its obstacle.
The check at 50
The first 70 stacks gave 2,822 layouts. None has a consistent fill at 50: 12 have three dead slots, none has four, and the other near layouts have five or more. Then, allowing a limited number of entries below 50 on the 1,404 layouts with at most eight dead slots, 1,317 proved inconsistent, and no timeouts occurred.
Entries allowed below 50
New fills
3 or 4
0
5
2
6
16
7
8
Two search faults
The search program stops at 50 stacks by default yet still reports the slice as complete. Slice 0 reached that limit at first row 3,870 of 7,451. I removed the limit and resumed from that row, so nothing was lost. Earlier rounds never reached 50 stacks in a slice, so they are unaffected.
On the night of 4 November the search then stopped at first row 6,420, because one GPU batch produced more results than its buffer holds. It sat idle for about 7.5 hours before I restarted it this morning with batches half the size, and it has passed that point. Progress is recorded only after a batch finishes, so no results were lost. At the last measured speed of about 210 to 290 first rows an hour, the rest takes about 5 to 7 days.
Guided compounds: a correction
I tested the split method by hiding 360 real entries and seeing how many it could recover. My earlier diagnosis, that the method was missing entries because of how it split patterns, was largely wrong.
Version
Hidden entries found (of 360)
STAND-UP COMEDY BIT (of 3 tries)
Original
228
1
Original, parts of 3 letters allowed
240
3
New prompt with ranked anchor words, no word lists
172
3
New prompt with ranked anchor words and word lists
240
3
On the real dead patterns, the best version offered 15 fitting entries from 147 usable splits, and none was new and usable: IRRATIONALIZING, SURGEON-DENTISTS, GLARELESS and INTORTIONS were already in the lexicon at 40, and WOODEN RAILS and BABY MARMOSETS failed the frequency screen. Since the method recovers about two-thirds of hidden real entries, my best guess is that these patterns have few well-known entries to find, which would make changing the stacks more promising than proposing more words for them. That is an inference, not a measurement.
Grid 9 is also in round40_grid9.pdf (grid, clues and answers) and round40_grid9.puz, in the “Claude Artifacts / crossword-stack” folder on Chris's Google Drive.
A mirror-symmetry grid with every entry at 50 or more: grid 8, Q 1.820
Eric graded the five canonical words that the next nearest grids needed, and all five passed: IRRATIONALITIES, INDIAN POTATOES and CORN FLEA BEETLE at 5.0, BINOMEN and AIN at 4.0, all with high confidence. With them, all 16 layouts of one mirror-symmetry stack fill with every entry at 50 or more. The best is grid 8 below.
1.820Q of grid 8, left-right mirror symmetry
16layouts of this stack that now fill at 50
5words graded by Eric for it, all 4.0 or 5.0
2grids at the 50 standard: grid 7 (up-down) and grid 8 (mirror)
Eric ranks the symmetries: full rotational symmetry is highly preferred, left-right mirror symmetry is worth exploring, and up-down symmetry alone is not. So grid 8 outranks grid 7 although its Q is a shade lower, and the search now concentrates on rotational and mirror layouts.
Grid 8 · Q 1.820 · mirror symmetry
Stack rows: THETHINBLUELINE, LABRETPIERCINGS, IRRATIONALITIES, SEASONTOBELIEVE and TOTHESAMEDEGREE, all graded 5.0 by Eric. INDIAN POTATOES and CORN FLEA BEETLE run down through the stack.
Entry
Eric's grade
Score
IRRATIONALITIESstack
5.0 high
90
INDIAN POTATOES
5.0 high
90
CORN FLEA BEETLE
5.0 high
90
BINOMEN
4.0 high
70
AIN
4.0 high
70
Sixteen entries score exactly 50, all list entries Eric has not graded, among them AYEM, SYCE, HAEM, CECILE, ANCON and RECELEBRATE. Two things an editor may object to: SHIT LIST (15 Down, on the lists at 80) and the bottom row, which has nine black squares in a row.
Eric also proposed a new kind of Borgmann step, “guided compounds”: split a dead slot's pattern in two, take real words that fit one part, and ask for completions that fit the rest, nearest meanings first (his example: STAND-UP COMEDY + BIT). It is built as guidesplit. Its first runs, on the 27 commonest dead patterns at 50 that split this way, gave nothing usable from any of six cheap models: the answers were listed halves glued together. Those patterns may have no good answer at all.
Grid 8 is also in round40_grid8.pdf (grid, clues and answers) and round40_grid8.puz, in the “Claude Artifacts / crossword-stack” folder on Chris's Google Drive.
The first grid with every entry at 50 or more: grid 7, Q 1.823 (up-down symmetry)
Chris has set the bar at 50, Eric's 3 (“usable under duress”), for every entry in the grid. A fresh band search at 50 was already done on 24 October and found no layouts in any symmetry, so I re-checked round 40's 7,996 layouts against the vocabulary at 50. The nearest had four dead slots. Grid 6 was one of them, and its four were exactly the canonical words we had scored 40 because no list grades them. Eric graded all four as usable, and the refill at 50 completes the grid.
1.823Q of grid 7, up from 1.808 for grid 6
50lowest entry score in the grid
4entries graded by Eric for this grid, all usable
107round-40 layouts within five dead slots of the 50 bar
Grid 7 · Q 1.823 · up-down symmetry
The same stack as grid 6. Two entries changed: 1 Across ICU became INU and 2 Down CAGS became NAGS.
Entry
Eric's grade
Score
SANTA MARIA TREESstack
4.0 high
70
TASAR
3.9 medium
68
SNYES
3.9
68
DEIRIDS
3.4 medium
58
His notes: SANTA MARIA TREES is “certainly a usable entry”; TASAR (a racing boat and a silk) is “usable for our purposes here”; SNYES is “just shy of publishable,” from French chenaille, channel; DEIRIDS is a dictionary entry in modern use, just obscure. His grade on the 0–5 scale maps onto our 0–100 bands (3 = 50–70, 4 = 70–90). Fifteen entries score exactly 50; they are list entries Eric has not graded, among them ASSHE, ISACT, ELD, OMEI, RNAS, SIG and NAGS. The other 32 layouts with four dead slots share one mirror stack (IRRATIONALITIES), and they stay dead even with their four canonical words admitted at 50.
Grid 7 is also in round40_grid7.pdf (grid, clues and answers) and round40_grid7.puz, in the “Claude Artifacts / crossword-stack” folder on Chris's Google Drive.
Round 40 ends, with a new best at Q 1.808 (up-down symmetry)
All 18 slices of round 40 have finished, and the final pass filled every layout from all of them. The six up-down slices ran much faster than the others, and the last of them produced the best grid of the round: grid 6 below, at Q 1.808, with only five entries below 50. It beats grid 4 (1.805, ten entries below 50) on both counts.
5entries below 50 in grid 6, all canonical words at 40
By symmetry the 751 fills are 435 rotational, 289 mirror and 27 up-down. The final pass filled four fewer rotational layouts than the two earlier partial fills (439); I have not yet found why. Round 40 limited the second stack row to 200 choices per stack, so it is a sample, not an exhaustive search.
Grid 6 · Q 1.808 · up-down symmetry
Stack rows: SANTAMARIATREES 40, TRAINEDASSASSIN 96, ACHRISTMASSTORY 80, CHOOCHOOCHARLIE 50, APOLLOASTEROIDS 50. Three more 15s run down: INDIANA HOOSIERS, THE ROAD TO ATHENS and DOUBLE SOLITAIRE.
Entry
Meaning
Counts
SANTA MARIA TREESstack
plural of Santa Maria tree, Calophyllum brasiliense, a tropical American timber tree
– / – / –
CAGS
plural of cag, British informal for a cagoule (in OED text)
29 / 68,418 / 44,382
DEIRIDS
plural of deirid, a sensory papilla near the head of a nematode
– / 518 / 3,774
SNYES
plural of snye, a side channel of a river (Canada)
– / – / 119
TASAR
a variant of tussore, the coarse silk of the tussah moth (Chambers, OED)
27 / 23,500 / 12,477
The rarest are SNYES and DEIRIDS. TASAR and SNYES are short downs inside the stack, and DEIRIDS crosses it. CAGS sits in the top-left corner, clear of the stack.
Grid 6 is also in round40_grid6.pdf (grid, clues and answers) and round40_grid6.puz, in the “Claude Artifacts / crossword-stack” folder on Chris's Google Drive.
289 fills under mirror symmetry (round 40, mirror slices 2–5)
Round 40 has produced its first fills with left-right mirror symmetry. Mirror slices 2 to 5 gave 2,591 layouts. 297 of them have a word that fits every slot, and 289 of those filled: 64 from slice 2, none from slice 3, 14 from slice 4 and 211 from slice 5. The first mirror slice had given 132 layouts and no fills. The best mirror fill, grid 5 below, scores Q 1.797, just under grid 4 (1.805), which stays the best grid of the round.
2,591new symmetric layouts (mirror slices 2–5)
289filled, of 297 consistent layouts
1.797best mirror Q (best overall is still 1.805)
7 of 18slices remain: the last mirror slice and up-down symmetry
Round 40 now has 728 complete fills: 439 with rotational symmetry and 289 with mirror symmetry.
Grid 5 · Q 1.797 · mirror symmetry
Stack rows: ARMISTEADMAUPIN 50, DOUBLESTROLLERS 86, EDRIOASTEROIDEA 40, DEADTREEEDITION 80, ADDISONSDISEASE 40. SSW (7 Across) is untinted: it is on the lists at 65, lowered to 45 by the initialism rule.
Entry
Meaning
Counts
EDRIOASTEROIDEAstack
an extinct class of disc-shaped Paleozoic echinoderms
– / 257 / 7,045
ADDISON'S DISEASEstack
chronic adrenal insufficiency, named for the London physician Thomas Addison; a common medical term, missing from the scored lists
– / – / –
BIDI-BIDIS
plural of bidi-bidi (biddy-biddy), the New Zealand burr plant Acaena
– / – / –
DIVI
British slang for a dividend (also divvy)
150 / 327,249 / 1,251,865
III
the Roman numeral three; all vowels, so Eric's rule ranks it low
68,030 / 58,343,659 / 244,843,876
MORDISHEEN
an old name for cholera (from Portuguese mordexim); an OED variant form
– / – / –
NATTES
plural of natte, a plait (French; in OED text)
– / 3,314 / 5,346
NIDULITES
Nidulites, a Paleozoic fossil genus
– / – / 5,521
ALOISIITE
a mineral name in Webster's Second; I could not confirm its composition
– / – / –
TEAR OFFING
on the canonical list in this spaced form; I can't find a sense beyond the verb tear off
– / – / –
The weakest are TEAR OFFING, ALOISIITE and MORDISHEEN. All three are downs that run through the stack, as are BIDI-BIDIS, NIDULITES and NATTES.
Grid 5 is also in round40_grid5.pdf (grid, clues and answers) and round40_grid5.puz, in the “Claude Artifacts / crossword-stack” folder on Chris's Google Drive.
192 more fills, and a new best at Q 1.805 (round 40, rotational slices 2–6)
The rest of the rotational part of round 40 has finished. Slices 2 to 6 gave 1,733 more layouts. 192 of them have a word that fits every slot, and every one of those 192 filled. The fills come from five stacks; one of them accounts for 142. The best, grid 4 below, scores Q 1.805, just above grid 1 (1.802). Every round now uses Eric's rule against all-vowel and all-consonant initialisms.
12 of 18slices remain: mirror and up-down symmetry
Across the 192 fills, 157 distinct entries score below 50. 145 of them are canonical words at 40. Eleven are initialisms on the scored lists that the new rule lowered to 45 or 41, among them GST, SSN and RPGS. ADAMA is on the lists at 49. Two are on neither list: POHAI and SSSSSSS. 151 of the 192 fills avoid both, grid 4 among them. None of these entries has been judged yet.
Grid 4 · Q 1.805
Stack rows: SACROSPINALISES 40, ASLINTERPRETERS 85, THESHARPERIMAGE 62, AERIALFUNICULAR 40, TRUSTLESSNESSES 40. GST (39 Across) is untinted: it is on the lists at 65, lowered to 45 by the initialism rule.
Entry
Meaning
Counts
SACROSPINALISESstack
plural of sacrospinalis, the muscle along the spine (erector spinae); a formal Latin plural
– / – / –
AERIAL FUNICULARstack
a cable-car railway, an aerial ropeway
– / – / –
TRUSTLESSNESSESstack
plural of trustlessness, lack of trust; the plural is formal only
– / – / –
CLERUCHES
Athenian citizens settled in a cleruchy abroad
– / – / –
H-SCOPE
a radar display showing a target's bearing and elevation
– / 274 / –
LETTERGAE
Scots, old: a church precentor (Chambers, OED)
– / – / –
PERFET
obsolete spelling of perfect (Spenser)
8 / 6,543 / 17,635
SCIRPUSES
plural of Scirpus, the bulrushes
– / – / –
The weakest are LETTERGAE and PERFET, both archaic. Both are downs that run through the stack, as do CLERUCHES and SCIRPUSES.
Grid 4 is also in round40_grid4.pdf (grid, clues and answers) and round40_grid4.puz, in the “Claude Artifacts / crossword-stack” folder on Chris's Google Drive.
Five-deep symmetric stacks from a larger lexicon (round 40)
The search produced 247 complete fills of symmetric 15×15 grids with a five-deep stack of 15s. All of them have 180° rotational symmetry and no 2×2 black blocks. Every entry is on the scored crossword lists or on a canonical word list. The three best are below, each from a different stack.
What found them. These fills are not the result of a true Borgmann search. They come from a larger lexicon and a search that no longer restricts it arbitrarily: every canonical word not on the crossword lists was admitted at score 40. No LLM proposed or judged any entry.
The search program has since been changed back to true Borgmann searches, in which the LLM proposes real words for each bottleneck pattern and the web check and Eric only verify those proposals.
0layouts on the scored lists alone (25,565 fifteens)
44,710fifteens with the canonical tier at 40
1,024symmetric layouts from 11 stacks (slice 1 of 18)
247of 261 consistent layouts filled
Update: Eric's edit of grid 1
Eric Chaikin called grid 1 “a very good grid” and, after a few hand changes to the two bottom corners, publishable. The black squares and the five stack rows are unchanged. His edit removes the canonical-only SAB and three entries that score 50 (TADDEO, SETTINGHEN and GCT). In their place it uses SETTING SUN, SADDEN, RAD, YEA, DOURLY and DUA. The one entry on neither list is GCS (Glasgow Coma Scale). On the scored lists it counts as 0, which puts the grid at Q 1.778. At a score of 50 it would be about 1.80.
Eric's edit. Tinted squares are the seven canonical-only entries that remain. GCS (51 Down) is untinted but is on neither list.
Slot
Was
Now
23 Down
SETTING HEN 50
SETTING SUN 50
42 Across
ETH 60
ETS 60
46 Across
DOESNT 70
DOURLY 50
47 Down
SAB 40
RAD 70
48 Down
NSA 65
LSU 60
49 Down
TEA 80
YEA 80
51 Down
GCT 50
GCS –
56 Down
SSO 60
SSN 65
59 Across
BAA 70
DUA 55
60 Across
TADDEO 50
SADDEN 75
Eric also asked to deprioritize abbreviations and initialisms made only of vowels or only of consonants. From now on every round's vocabulary takes 20 points off each such entry before the score cut. That covers 1,140 entries, among them WGN, TSPS, EEO, SSN and GCT. A Y exempts an entry either way, and so do real words such as BRR, HMM, TSK and EAU. The new rule would have demoted SSN in Eric's edit too.
The 331 entries scored below 50 in those fills were checked against the canonical list: 325 are canonical and 6 are not (XEN, NENAD, OOAA, HSIFAN, ENGINESIZING, GOTTENSWEPTUPIN). 219 fills avoid all six, including all three grids here. In the grids, tinted squares belong to an entry that is canonical but not on the scored lists. Q is the quality index. Counts are Usenet / Google Web 2006 / Google Books 2020; a dash means none recorded.
plural of alcoholisation (British): saturation with alcohol
– / – / –
SPOTTAIL SHINERSstack
the North American minnow Notropis hudsonius
– / – / –
ANAMNIOTE
vertebrate without an amnion (fish, amphibians)
– / 251 / 3,144
CENTGENER
plant breeding: a plot of 100+ plants from one parent
– / – / 11,967
HIBA
the Japanese conifer Thujopsis dolabrata
21 / 26,398 / 35,018
ILIANS
inhabitants of Ilium, Trojans
4 / 218 / 8,639
PILE PLANK
a plank driven as sheet piling
– / – / –
Each grid also exists as a PNG and as an Across Lite .puz file with a clue for every entry, for opening in Crossfire. Eric's edit is round40_grid1_eric.puz. Those files are in the “Claude Artifacts / crossword-stack” folder on Chris's Google Drive.