11-11-2026, 12:20 AM
I found something that my fake Voynich generator (the skribe, no meaning inside) doesnt do, and the real text does.
Every hand has its usual words. On some pages it uses the most common ones a lot, on other pages it goes more for the less common ones. Same vocabulary, just a different mix. I call it the “temperature” of the page.
The funny thing is that it goes by bifolio. The four pages of a bifolio have the same temperature, even when the two leaves end up far apart in the book. Two bifolios of the same quire are not similiar at all. Pages of the same bifolio correlate around 0.4, in the generator around 0.05.
Its not the transcription, I get the same with Takahashi, Glen Claston and Zandbergen/Fagin Davis. Pages where the transcribers were unsure are not hotter or colder, so its not misreadings. And it doesnt slowly change from the start of the book to the end, it just jumps from one bifolio to the next.
If I give the generator one small random push per bifolio (a bit more common words, or a bit more rare ones) it does the same thing, and still passes the other tests.
So it looks like each bifolio was written in one go, before binding, with its own pace or mood. Davis & Layfield saw something similar with LSA. It could also be one topic per bifolio, I cant tell the two apart yet.
A word about the skribe, becuase people sometimes think it learns the language. It doesnt. It has no meaning, no sentences and no codebook. It only copies writing habits from the real text: which glyph tends to follow which (counted per hand and section), some rules for line starts, line ends, first lines and drawings, and now a random "mood" per page and per bifolio. Then it writes glyph by glyph. So when it matches something, that something can be explained by habits alone.
This is where it is now, all 88 tests I use, real against the scribe (average of 5 texts). A test passes if the scribe's number is less than 25% away from the real one, so for example if the real text gives 4.0 the scribe has to land between 3.0 and 5.0 (for the tiny ones near zero, like 10 and 11, it just has to be within 0.01). It passes 85, before the bifolio mood it was 84. Line 89 is the new one.
The three that fail: frequent words that avoid each other (the real text has fewer of them), -m at the end of long lines against short ones (the real effect is stronger), and q- on the first word of a paragraph (a bit too many in the scribe). Everything else is inside 25%.
Every hand has its usual words. On some pages it uses the most common ones a lot, on other pages it goes more for the less common ones. Same vocabulary, just a different mix. I call it the “temperature” of the page.
The funny thing is that it goes by bifolio. The four pages of a bifolio have the same temperature, even when the two leaves end up far apart in the book. Two bifolios of the same quire are not similiar at all. Pages of the same bifolio correlate around 0.4, in the generator around 0.05.
Its not the transcription, I get the same with Takahashi, Glen Claston and Zandbergen/Fagin Davis. Pages where the transcribers were unsure are not hotter or colder, so its not misreadings. And it doesnt slowly change from the start of the book to the end, it just jumps from one bifolio to the next.
If I give the generator one small random push per bifolio (a bit more common words, or a bit more rare ones) it does the same thing, and still passes the other tests.
So it looks like each bifolio was written in one go, before binding, with its own pace or mood. Davis & Layfield saw something similar with LSA. It could also be one topic per bifolio, I cant tell the two apart yet.
A word about the skribe, becuase people sometimes think it learns the language. It doesnt. It has no meaning, no sentences and no codebook. It only copies writing habits from the real text: which glyph tends to follow which (counted per hand and section), some rules for line starts, line ends, first lines and drawings, and now a random "mood" per page and per bifolio. Then it writes glyph by glyph. So when it matches something, that something can be explained by habits alone.
This is where it is now, all 88 tests I use, real against the scribe (average of 5 texts). A test passes if the scribe's number is less than 25% away from the real one, so for example if the real text gives 4.0 the scribe has to land between 3.0 and 5.0 (for the tiny ones near zero, like 10 and 11, it just has to be within 0.01). It passes 85, before the bifolio mood it was 84. Line 89 is the new one.
| # | test | real | scribe | pass? |
| 1 | Adjacent identical words (%) | 0.90 | 0.94 | ok |
| 2 | Adjacent identical words / chance | 3.09 | 3.30 | ok |
| 3 | Consecutive words at edit distance 1 / shuffled | 2.21 | 1.76 | ok |
| 4 | Same word at distance 2 / shuffled | 3.27 | 3.74 | ok |
| 5 | Same word at distance 3 / shuffled | 3.45 | 3.52 | ok |
| 6 | Same word at distance 5 / shuffled | 2.85 | 3.43 | ok |
| 7 | Frequent words repeated >3x chance (%) | 51.8 | 42.2 | ok |
| 8 | Frequent words repeated <0.3x chance (%) | 16.1 | 25.5 | FAIL |
| 9 | Boundary MI, distance 1 (last glyph to next first glyph) | 0.15 | 0.15 | ok |
| 10 | Boundary MI, distance 2 | 0.009 | 0.004 | ok |
| 11 | Boundary MI, distance 3 | 0.003 | 0.003 | ok |
| 12 | Lift -y | q- | 1.69 | 1.66 | ok |
| 13 | Lift -r | a- | 3.12 | 3.14 | ok |
| 14 | Lift -y | a- | 0.19 | 0.19 | ok |
| 15 | Lift -n | q- | 0.47 | 0.53 | ok |
| 16 | First word of line starts ch/sh (%) | 8.15 | 8.29 | ok |
| 17 | First word of line starts y (%) | 14.3 | 15.0 | ok |
| 18 | Last word of line ends -m (%) | 15.9 | 19.0 | ok |
| 19 | Line-end -m: lines of 10+ words / 1-4 words | 3.40 | 2.49 | FAIL |
| 20 | Same first glyph as line n-1 / null | 0.58 | 0.59 | ok |
| 21 | Same first glyph as line n-2 / null | 1.30 | 1.29 | ok |
| 22 | Same first glyph as line n-3 / null | 1.16 | 1.12 | ok |
| 23 | Distinct initial glyphs per paragraph (4+ lines) | 4.36 | 4.51 | ok |
| 24 | Adjacent repeats: first line of paragraph / rest | 0.38 | 0.41 | ok |
| 25 | Spearman correlation of neighbour word lengths | 0.16 | 0.20 | ok |
| 26 | Word types | 7,915 | 7,606 | ok |
| 27 | Hapax (% of types) | 71.3 | 65.7 | ok |
| 28 | Words fitting the slot grammar (%) | 97.7 | 99.9 | ok |
| 29 | Words with 2+ gallows (%) | 1.80 | 1.89 | ok |
| 30 | Mean word length (glyphs) | 4.80 | 4.79 | ok |
| 31 | SD of word length | 1.71 | 1.76 | ok |
| 32 | e and i in the same word / expected | 0.24 | 0.24 | ok |
| 33 | H(prefix), marginal | 1.54 | 1.52 | ok |
| 34 | H(prefix | previous glyph, core) | 0.92 | 0.93 | ok |
| 35 | q- before o- after -y (%) | 57.3 | 58.7 | ok |
| 36 | q- on first word of paragraph (%) | 4.19 | 5.30 | FAIL |
| 37 | q- on first word of line (%) | 14.9 | 16.2 | ok |
| 38 | q- inside the line (%) | 17.9 | 19.0 | ok |
| 39 | a- at line start (%) | 0.35 | 0.32 | ok |
| 40 | d-/s-/y- at line start (%) | 48.0 | 46.9 | ok |
| 41 | Bare gallows at line start (%) | 8.97 | 9.65 | ok |
| 42 | Closed last word, non-final lines (%) | 27.1 | 29.1 | ok |
| 43 | Closed forms inside the line (%) | 2.34 | 2.33 | ok |
| 44 | Closed last word of paragraph (%) | 17.3 | 17.6 | ok |
| 45 | Length of 2nd word / 1st word | 0.88 | 0.92 | ok |
| 46 | Gallows on first word of paragraph (%) | 93.2 | 92.5 | ok |
| 47 | p/f among gallows, first line of paragraph (%) | 31.0 | 29.8 | ok |
| 48 | p/f among gallows, other lines (%) | 2.73 | 2.82 | ok |
| 49 | Plain p/f followed by e (%) | 0.31 | 0.33 | ok |
| 50 | Plain p/f followed by ch/sh (%) | 59.2 | 57.5 | ok |
| 51 | Word length, first line / rest | 1.08 | 1.01 | ok |
| 52 | Closed word before a drawing gap (%) | 10.2 | 10.1 | ok |
| 53 | d-/s-/y- after a drawing gap (%) | 41.3 | 41.5 | ok |
| 54 | q- after a drawing gap (%) | 3.57 | 4.08 | ok |
| 55 | Same first glyph after a gallows-initial line / null | 0.61 | 0.68 | ok |
| 56 | Words with an edit-distance-1 twin earlier on the page (%) | 64.5 | 65.7 | ok |
| 57 | Lines back to the nearest twin (median) | 2.00 | 2.00 | ok |
| 58 | Page overdispersion of q- (after -y) | 6.32 | 6.31 | ok |
| 59 | Vocabulary cosine: same sheet / neighbour pages, other sheets | 1.30 | 1.48 | ok |
| 60 | Strange words (rare trigram in the real text) (%) | 4.59 | 3.70 | ok |
| 61 | Strange words per line | 0.36 | 0.29 | ok |
| 62 | Page overdispersion of e-words | 24.6 | 24.9 | ok |
| 63 | e-words: biological / herbal | 2.04 | 1.90 | ok |
| 64 | -edy words in biological (%) | 27.5 | 29.4 | ok |
| 65 | Share of the page's most frequent word (median, %) | 4.71 | 4.61 | ok |
| 66 | Words of 8+ glyphs (%) | 5.25 | 6.13 | ok |
| 67 | Character entropy h1 | 3.91 | 3.90 | ok |
| 68 | Conditional entropy h2 | 2.14 | 2.10 | ok |
| 69 | Conditional entropy h3 | 1.86 | 1.83 | ok |
| 70 | Zipf slope (first 1,000 types) | 1.00 | 0.98 | ok |
| 71 | Words of 2-3 glyphs (%) | 21.6 | 22.7 | ok |
| 72 | Words of 5-6 glyphs (%) | 41.0 | 39.1 | ok |
| 73 | Shared vocabulary, neighbouring pages (cosine) | 0.33 | 0.30 | ok |
| 74 | Shared vocabulary, 10 pages apart | 0.25 | 0.22 | ok |
| 75 | Shared vocabulary, 100 pages apart | 0.15 | 0.15 | ok |
| 76 | Neighbouring / 100 pages apart | 2.13 | 1.95 | ok |
| 77 | Bursts of the 50 most frequent words (CV of gaps) | 1.75 | 1.69 | ok |
| 78 | Page overdispersion of the 50 most frequent words | 2.06 | 1.97 | ok |
| 79 | Words starting with a gallows, not line-initial (%) | 5.05 | 5.01 | ok |
| 80 | Gallows-initial word after another / chance (in line) | 1.45 | 1.44 | ok |
| 81 | sh share (sh/ch), first line / other lines | 1.34 | 1.31 | ok |
| 82 | t share (t/k), first line / other lines | 1.23 | 1.27 | ok |
| 83 | frequent words, first line / other lines | 0.66 | 0.67 | ok |
| 84 | words with ee, first line / other lines | 0.87 | 0.86 | ok |
| 85 | Line-final -m written -am, not -im (%) | 94.5 | 96.0 | ok |
| 86 | Plain p among plain p+f, first line (%) | 75.4 | 78.3 | ok |
| 87 | q- after -y before o, last line / other lines | 0.81 | 0.80 | ok |
| 88 | ch- words, last line / other lines | 1.25 | 1.23 | ok |
| 89 | Same bifolio, page to page correlation of temperature | 0.39 | 0.36 | ok |
The three that fail: frequent words that avoid each other (the real text has fewer of them), -m at the end of long lines against short ones (the real effect is stronger), and q- on the first word of a paragraph (a bit too many in the scribe). Everything else is inside 25%.


![[+]](https://www.voynich.ninja/images/collapse_collapsed.png)