9x9 Column Proposals

9x9 Column Proposals

Borgmann search, round 4. For Chris and Eric. 6 Nov 2026. Written by Claude.

Summary

Clean scan: first real six-row partials

Without the fill tier, every row and column comes from cat1 plus words I reviewed. The search is then small enough to nearly finish. With rows 1–3 limited to common words, it covers all 5,375 start words in about two minutes, and 24 are cut by budget.

Scratch lexiconTop rowsFive-rowSix-rowSeven-rowSquares
1,128 wordscommon76,300800
+ 175 wordscommon76,9952600
+ 9 wordscommon76,9983500
+ 20 reviewed fill wordscommon77,0733500
+ 65 words (1,397 in all)common77,1863500
Before those 65 (1,332 words)wider875,10311600
+ 65 words (1,397 in all)wider876,50220000
Larger reviewed lexicon (lex12)wider1,112,41125200
+ 159 unaimed field words (lex13)wider1,129,45925300
Judged fill tier (lex15, 11,238 words)common241,6785200
Judged fill tier (lex15, 11,238 words)wider3,078,45943000

Wider means 13,193 start words. Cuts by budget vary by run: about 1,890 in the first wider runs, 5 in lex13, 1,635 of 13,193 in the wider lex15 run, so that row is a sample and is not directly comparable with lex13. The fill-tier rows use strings accepted by judges, most of them not real words by my reading. The other added words are my reading, not graded by Eric.

One example: sharpness / tolerance / rotterdam / overscore / merosomes / brightest, which relies on my pressheet.

Row 7 is the new wall. Of 123 six-row partials checked in the wider scan, only 2 have a listed row word within two column misses. No real word is in sight for prefixes such as tinamib or flaptrf.

What free columns buy

This section and the next use the older setup with the fill tier. I took the 3,272 five-row partials from its last scan and resumed them.

Rows came from the lexicon. Columns could also use strings from the 9-letter corpus pool, which is full of junk, so the two pool rows of this table are a ceiling.

Columns may useSix-rowSeven-rowEight-rowSquares
Lexicon only0000
+ pool, count ≥ 100 (1.36M)8,9271300
+ all pool (2.48M)59,94227700
+ 71 real words I picked0000
+ those 71 and 687 judge-accepted strings70100

Fresh 48-start scans: with the 71 words, 3,255 lines at five rows and 0 at six. With the judged strings too, 3,305 at five, 68 at six, 1 at seven. The base was 3,243 at five. All starts were cut by budget, so these are samples. Candidates were chosen because they fit these partials, so some of the gain is by construction.

Why real words fall short

One of the easiest six-row partials. It needs two new column words, starting restop and sealss. The first could be real; the second is hard to believe.

Few partials are this easy, and each new word also pins the letters of the rows below it.

New column words needed (pool count ≥ 100)Six-row partials
11
234
3211
4 or more8,669

Counted from up to three partials stored per pattern. Every seven-row partial in this group needs five or more.

The strings these partials lean on are mostly junk or names: seanessie, stiverson, sellstrom. A few are real: carnaroli, seastacks, stoplists.

llama3.1:8b could not propose from prefixes: 220 prefixes gave 0 usable words before the run was stopped.

The six-row partials closest to real, by my reading:

The one seven-row partial needs at least appalatus or appalatia, saliaceae and seaness…, none of them real words.

Judges

Scored from verdicts already on file. Each cell is strings accepted. A second judge sees only what gemma4 rejected, about 29% of strings.

Acceptedgemma4 alone+ gemini-2.5-flash+ mistral-large-2512
Real words (of 200)172184182
Junk (of 72)222722
Eric ≥ 3 (of 142)112128119
Eric < 3 (of 39)142420

On the 1,675 column candidates, gemma4 accepted 544 and took 610 s. Gemini then accepted 143 of the 1,131 strings left, for $0.04 (46 calls, none failed). Both lists hold much junk, such as othermeat and seanessie. The test sets above were cleaner than these candidates, so don't expect those rates here.

Mistral failure

mistral-large-4-0 now refuses requests with reasoning off (HTTP 400, code 3051), which is what word_oracle sends. With reasoning on it answers, at $1.14 per 1,000 strings and two minutes per batch. word_oracle now uses mistral-large-2512 as its second judge. A single probe call passed; it has not been run end to end.

Limits of these results

Decisions needed