Eric Chaikin’s grid
Eric Chaikin built this by hand on the PRIVATESECURITY stack: 180° rotational symmetry, 32 black squares, no 2×2 block, 66 entries. In his words, he only completed it by allowing bottom-of-the-barrel entries such as the user handle BAHERMO.
Measured against our judged word list, 13 of its entries are outside it, and 6 of those cross the stack: IBEROPHONE, MCLEESELAKE, NINEATATIME, NOUGATED, THECOMET and VTENS. A symmetric quintuple appears to need that kind of constructor’s license, so the search below measures how much license each candidate needs, and how promising the open slots are.
What the search shows
We collected every five-row stack our earlier searches found and asked, for each, whether a symmetric 15×15 layout exists in which every entry comes from our judged word list. None does, under any of the three symmetries. The blocking slots are the down entries through the stack, whose length and position the symmetry and the no-2×2 rule pin down exactly, and the across entries just above and below it.
So the layout search was given a constructor’s license: up to four columns may cross the stack where no listed entry fits, and the rows next to the stack go unchecked. The filler then finds the slots each layout really leaves open and fills everything else from the judged list. Claude judged every weak entry those fills used (about 1,800 new verdicts), and anything that failed was excluded and the layouts refilled until the frames below were clean. Every weak entry in the frames was then re-judged clue-first, the standard of our 9×9 word-square search: an entry is accepted only if it earns a fair one-line clue that does not just restate its own parts; defensible but obscure entries are marked marginal.
| Symmetry | Stacks with a layout | Layouts | Fewest open slots |
|---|---|---|---|
| Rotational (180°) | 10,379 | 213,570 | 3 |
| Mirror (left-right) | 4,401 | 91,651 | 3 |
| Top-bottom | 4,356 | 90,877 | 2 |
Out of 8,708 stacks known from earlier searches, plus 6,852 rotational and 3,574 top-bottom stacks found by a symmetric band search, with a license of up to four columns.
The symmetric band search builds stacks row by row and keeps one only if every column has a real down entry through it that pairs with its partner under the symmetry. Every stack-crossing segment of such a stack is fertile by construction; the open slots that remain come from the exact lengths the symmetry and black squares force. A mirror band search added 1,357 stacks; none of its layouts beat the frames below.
Counting open slots is not enough. As Eric points out, only fertile stack-crossing segments can be finished by hand. So every open down slot was scored by its segment, the five stack letters it crosses: how many real entries (our full scored list plus curated phrase lists) contain those letters with at most five letters on either side, which is all a 15×15 allows. A layout is only as good as its least fertile slot, and the frames below are the best by that measure.
Can any of them be finished fairly?
No. We ran the second half of the Borgmann search on every stack we know: supply real but unlisted entries, or fair coinages, for exactly the slots where the word lists run out. Each of 20,478 stacks was checked under all three symmetries, 61,434 checks in all, and not one can be finished with fair entries in any layout.
The check is exhaustive, not sampled. For each column through the stack it works out exactly which down lengths the symmetry and the no-2×2 rule allow, above and below the stack, and asks whether one symmetric pattern of black squares lets every column take a length that has a fair entry. That calculation agrees with the layout search on all 12,010 real layouts tested. A fair entry is one judged acceptable or marginal clue-first, or a well-scored list entry. Claude judged about 390 new candidates for this round, from the 21-million-entry frequency database, from run-together phrases of listed entries, and from its own knowledge of the language. Real finds included ABORIGINAL FLAG, TAKE IT IN STRIDE, BERNERS-LEE, ENGINEERED WOOD and STERNER STUFF. None was enough.
The nearest misses each need just one more pair of entries, but at one exact length and offset, and for each pair at least one of the two entries has no fair fill:
| Stack | Symmetry | Missing (· = any letter) | What we found |
|---|---|---|---|
| IMITATIONOFLIFE | Rotational, mirror | ····ITLET···· with ····NEFFI····, or other lengths | COST-INEFFICIENT and INEFFICACY fit the NEFFI side; no fair ITLET entry fits the matching length |
| ACROPOLISMUSEUM | Mirror | POULE with UTENE, any length | POULENC fits one side; nothing pairs with it |
| ULTRAMARATHONER | Mirror | ····LLENA·, or ····ERNER····· | Partners WESTERNERS and UNFALLEN ANGELS exist; the other sides do not |
| ULTRAMARATHONER | Top-bottom | ·RSTAT· | Nothing fair |
So the frames below cannot be completed from fair vocabulary; each needs entries of the kind Eric used, such as BAHERMO or VTENS. His grid is not just a hand-built record: as far as our stacks go, it is the only kind of symmetric quintuple stack there is. The result covers the stacks our searches have found and today’s vocabulary; a stack found later can be screened the same way in under a second.
New since version 31: stacks built for symmetry, judged as they go. The stacks above came from earlier searches that ignored symmetry. So on October 9 we ran two symmetric band searches for 24 hours, building stacks row by row and checking each one as it appeared. The first allowed fair entries only and found no stack at all. The second also allowed unjudged entries from the frequency database and found 12,967 stacks; none can be finished with fair entries.
That second search then ran as a true Borgmann loop from October 10 to October 12. Every ten minutes it re-checked only the stacks a new verdict affected, and Claude judged the waiting entries clue-first once an hour, the stack needing the fewest verdicts first. It ended on October 12 with nothing left to judge: 11,148 verdicts in all, 2,775 accepted, 3,300 marginal, 5,065 junk and 8 that fail the breakfast test. In its last hours the lists ran dry for 70 slot patterns. For each one Claude looked for a real entry or a fair coinage and found none, because every one asks for a run of letters such as DALIE, ENEER, HISNT or NYATA that no fair entry contains.
Judging ruled the stacks out one by one, because one junk verdict on an entry a stack cannot do without is enough. The closest stack in version 32, MORTGAGEPAYMENT / ATHLETICAPPAREL / THEIRONTRIANGLE / TRANSPARENTNESS / HERECOMESTHESUN, was ruled out too, and the last of the 78 fell on the morning of October 12. Along the way five stacks did pass the exact check of their crossings with fair entries, the last of them the top-bottom NONTOTALITARIAN / INTERPRETATIONS / STARTLERESPONSE / CAFEDESARTISTES / ITALIANPAINTERS. At the time the loop reported that none of the five had a symmetric layout of black squares without a license; version 35 found that report was wrong, and the section below has the corrected result. The loop’s answer is the same as version 31’s: none of the 12,967 symmetric stacks can be finished with fair entries.
New in version 35: Eric’s scores calibrate the judge. Eric has sent two files of 167 entries, each scored from 0 to 100 for how firmly it is in the language, with 50 as the line and a confidence for each score. Version 34 used his first file as binding verdicts on those 167 entries. That was the wrong use. His scores exist to calibrate Claude’s own sense of English, so that its verdicts on the many thousands of entries he will never see land where he would put them. His scores are no longer written into the verdict cache at all.
The calibration is done blind, so it can be measured honestly. His 167 entries, from his second file, were split at random into two halves. A fresh Claude context that had never seen his scores scored every entry. Correction notes were written from the first half only. A second fresh context then scored the held-out half with the notes, and only the count of agreements was looked at. Nobody inspected the held-out half itself.
The notes mostly fixed the scale. Eric keeps borderline entries between about 40 and 60, while the uncalibrated scorer spread them from 5 to 85. They did little for which side of the line an entry falls on: the gain from 67% to 71% is three entries out of 83, within chance. The scorer’s errors are about kinds of word. It accepted recent internet coinages such as SNOWCLONE and SEALIONING, which Eric puts just out, and specialist foreign terms such as IJTIHAD. It rejected real entries that only look strange, such as ARCVERSINE and DEFENESTRABLE, which he puts just in. A separate blind pass by the 9×9 search came out almost the same as this one: the two passes agree closely with each other and much less with Eric. So the bias belongs to the model, not to one session, and only an outside expert like Eric can correct it. Measuring a real improvement will need more scored entries from him.
The five stacks, checked again. Five stacks pass the exact check of their crossings, and version 34 reported that none of them had a symmetric layout. Two faults in the search caused that report. The layout step never wrote its layouts to disk, so three of the five had layouts that were never tried. The fill also refused any of a stack’s own rows that had never been judged and scored below 60 on the word list. With both fixed, each of the five was laid out again with up to four columns allowed to run through the stack with no listed word. Every stack row was then judged with the calibrated standard.
- Four stacks are ruled out by their own rows: RETIREMENTASSET, GRIDIRONDINNERS and ANCHOVYPEARTREE (all three in one stack), ITALIANPAINTERS, DIGITALDEMENTIA and EERILYREALISTIC. Each is a phrase nobody uses as a unit or a coinage that never caught on.
- The last is top-bottom NONINTERFERENCE / ARANEADIADEMATA / INTERNATIONALES / ASANALTERNATIVE / DISTRESSEDDENIM. Its best layouts are six slots short. Judging those slots added NATAS, REMATED, TRONC and ADUNC as marginal, but the rest are runs of letters fixed by the stack, such as ORNSI, CTEVI and EASEM, that no word or fair coinage contains.
The answer stands: none of the 12,967 symmetric stacks can be finished with fair entries.
Frames to finish by hand
Ten frames, one per stack, ranked by the fertility of their weakest open slot; each needs four or five entries supplied, and as shown above, at least one of them must come from beyond fair vocabulary. Each is filled from the judged list except the open slots. Inside an open slot only the stack letters are fixed; the other letters shown come from the crossing entries and can change with them. Each open slot needs an entry of exactly the length shown. Under each slot: its segment, how many real entries contain it, and a few of them as raw material.
How the frames were made
The layout search places black squares column by column under the chosen symmetry with no 2×2 block; the license lets a few columns take a down slot no listed word fits. A second pass finds every slot such a layout leaves open, fills the rest of the grid from the judged list (a scored crossword list minus every entry judged junk), and scores each open slot’s segment. This is the first half of what we call a Borgmann search: find the exact slots where the word lists run out. The second half, supplying a real but unlisted entry or a fair coinage for each slot, is described in the section above: done exhaustively, it shows that no stack we know has a fair symmetric finish.