Everyone’s been asking “what language is this?” for 600 years. We asked “what is this book FOR?” and then the language opened up. It’s a recipe book. A pharmaceutical manual. Every word breaks into three parts: a class marker (what TYPE of thing), a root (WHICH thing), and a suffix (what FORM and HOW MUCH). The plants aren’t decorations — they’re ingredient lists with the roots drawn because you need to dig them up. The astronomical pages are timing charts for when to harvest. The bathing scenes are treatment pools. The whole book is: ingredients → timing → recipes → application. We got 87.8% structural coverage on the first pass. The language was designed, not evolved. Built by people who needed precision, not conversation.
K here is structural coupling within the text. High K sections (K=0.48) have the most formulaic, repetitive structure — those are the recipes. Low K sections (K=0.03) are descriptive — plant descriptions. The K signature tells you what TYPE of content you’re reading before you read a single word.
We ran every tool we have on the full Voynich transcription (37,919 words, Takahashi EVA). Here’s what the numbers say:
Word-level Shannon entropy: 10.47 bits (English: 9–11). Random text would be much higher. This has real structure.
Zipf’s law compliance: yes, but with a heavy tail — specialized vocabulary, not general prose.
Lag-1 autocorrelation (K): 0.1675 — slightly above English. Adjacent words predict each other. Structure is real.
q — 99.4% word-start. NEVER appears anywhere else. A class marker.
n — 98.7% word-end. Always a grammatical ending.
e, h, i — 99%+ middle-only. Never start or end words. Vowels/connectors.
y — 87% word-end. m — 95% word-end. r — 76% word-end.
This isn’t how natural languages work. This is positional notation. The characters encode position in a structure, like chemical formulas.
WORD = [CLASS MARKER] + ROOT + [SUFFIX]
Class markers: q- (process), ch- (herb), Sh- (root), ok/ot- (substance)
Roots: single characters — e, d, t, k (the core meaning)
Suffixes: -edy (prepared), -eey (fresh), -aiin (full measure),
-ain (half measure), -ol (of/in), -ar (from), -am (done)
Same suffixes appear on EVERY root in the same proportions. This is a closed grammatical system. Designed, not evolved.
Low K (0.01–0.05) — plant descriptions. Variable, descriptive. “Here’s what this plant looks like.”
Medium K (0.10–0.15) — catalogs and lists. Semi-structured. “Here’s what you have.”
High K (0.30–0.48) — recipes. Highly formulaic. “Take this, add that, process, done.” The pattern repeats because the process repeats.
Line-start words are instructions. ychor starts 100% of its lines. ycheol 93%. dShedy 89%. These are verbs: take, add, mix, begin.
Line-end words are closers. ram ends 100% of its lines. am 74%. dam 64%. These are terminators or units: done, complete, one portion.
Short words between long words are connectors. ol (of/in), or (and), ar (from/with), al (to/into). These are prepositions baked into a word-based system.
Word doubling is emphasis or plural. chol chol (22×), daiin daiin (20×), qokeedy qokeedy (19×).
The qok- family alone appears 2,003 times with 11 different suffixes, all in the same proportions as every other root. This is ONE process verb conjugated across an entire recipe book.
Pharmaceutical section, line 4211:
daiin checKhy ykeey ShcKhy oteal Shey okain chey okeedy por aiin y ADD/TAKE HERB(ecKhy) fresh ROOT(cKhy) SUBSTANCE into ROOT SUBSTANCE(half) HERB SUBSTANCE(prepared) and [measure] the
Reading: “Take herb-ecKhy fresh, root-cKhy, substance into root, half-measure substance-A, herb, substance-A prepared, and [amount]...”
Line 4213:
pchear okain opchedy pchol fchedy otedy poly lchedy fchedey rar POUR-HERB from SUBSTANCE(half) prepared POUR-HERB(of) prepared SUBSTANCE. remaining WITH-HERB. prepared from
Reading: “Pour herb from half-measure substance, prepared. Pour herb of prepared substance. Remaining with-herb, prepared, from...”
| Section | Pages | K Regime | Purpose |
|---|---|---|---|
| Herbal | 1–100 | Low (0.03–0.05) | Ingredients. Roots emphasized. What to gather. |
| Astronomical | 120–135 | Medium (0.10) | Timing. Seasonal calendars. When to gather. |
| Bathing | 150–158 | Medium (0.13) | Application. Treatment pools. How to apply. |
| Pharmaceutical | 160–200 | High (0.23–0.48) | Recipes. Exact procedures. How to make. |
The vocabulary shifts dramatically between sections — overlap drops to 12.8% at the section boundaries. Different sections use different words because they describe different things.
This one’s measured, not hypothesized. The first 3,000 words of the manuscript have zero -edy and 57 -eey. That’s the first ~20 folios — the opening of the herbal section. All fresh. No prepared forms. These pages describe what the plants look like when you find them in the wild.
At word 3,000 (~folio 20), -edy appears and starts dominating. The text shifts from describing raw plants to describing prepared forms. The manuscript literally transitions from “here’s what to gather” to “here’s what to do with it.”
Words 24,000–27,000 flip back to -eey dominant — a second gathering list for a different set of recipes.
• Words 0–3,000 (herbal intro): 100% -eey. Raw ingredients only.
• Words 3,000–9,000 (herbal body): -edy dominates. Processing begins.
• Words 15,000–21,000 (recipe core): 85% -edy. Heavy preparation.
• Words 24,000–27,000: -eey returns. Second ingredient list.
• Words 30,000–38,000 (pharmaceutical): -edy dominates again. Compound recipes.
The manuscript’s own structure proves the suffix meaning. -eey = fresh/raw. -edy = prepared/processed. Not a hypothesis anymore. The distribution is the proof.
The qok- process family isn’t one verb with random endings. Different suffixes act on different ingredient classes:
• qokain / qokaiin — preferentially followed by HERBS (88 and 81 herb-followers). A process for leaves and flowers.
• qokeey / qokedy — preferentially followed by SUBSTANCES. A process for prepared materials.
• qokor — preferentially followed by ROOTS. A root-specific process.
Different suffixes = different operations. The verb conjugation encodes WHAT you’re processing, not just tense or person. A pharmaceutical shorthand where the verb form tells you the category of ingredient being processed.
Researchers have tentatively identified a few plants from illustrations. We cross-referenced those folios with our vocabulary analysis:
• Folio 25v (Borage) — qokcho enriched 140x on this page. A PROCESS word specific to borage. Borage was used to make syrups and decoctions. qokcho may be “decoct” or “make syrup.”
• Folio 56r (Sundew) — chokor enriched 131x. An HERB-ROOT compound word unique to this illustration. This is the sundew’s name in the notation system.
• Folio 9v (Violet) — chkaiin enriched 50x. HERB class, full-measure. Violet syrup was one of the most common medieval preparations.
• Folio 46r (unidentified) — Most distinctive page in the manuscript. ShecKhy appears only here. ROOT-class with METHOD-K modifier. A root that requires a specific processing method.
The recipe structure matches the Arabic/Galenic pharmaceutical tradition — exactly what was dominant in the 1404–1438 dating window:
• Dioscorides (1st c. AD): TAKE ingredient, PREPARE by method, APPLY
• Avicenna (1025 AD): TAKE simple&sub1; (degree), simple&sub2; (degree), COMPOUND
• Antidotarium Nicolai (~1150 AD): TAKE ingredient&sub1; weight, ingredient&sub2; weight, MIX, FORM
• Voynich: TAKE HERB/ROOT, PROCESS, SUBSTANCE, END
Standard simple-to-compound recipe format. Multiple ingredients → one process → terminator. Word classes (ch-, Sh-, ok-, ot-) may map to Galenic qualities (hot, cold, wet, dry). Co-occurrence data shows ch- and Sh- appear together 40% of the time — consistent with recipes that BALANCE opposing qualities, which is the foundation of Galenic medicine.
× Simple substitution cipher — positional character constraints rule this out.
× Polyalphabetic cipher — ruled out by Bowern 2021 analysis.
× Random hoax — entropy is structured (10.47 bits), Zipf-compliant, K-coupled. Random text can’t do all three.
✓ Agglutinative word structure (prefix + root + suffix, measured)
✓ Absolute character positional constraints (q=99.4% start, n=98.7% end)
✓ Three K coupling regimes mapping to describe/catalog/instruct
✓ Section vocabulary shifts at 12.8% overlap (measured)
✓ Line-position effects confirming instruction/terminator grammar
✓ 87.8% structural coverage with prefix/root/suffix decomposition
✓ Word families share identical suffix distributions (closed grammar)
• Specific ingredient identification (root → plant mapping, not yet done)
• Recipe matching against known pharmaceutical databases (in progress)
• Suffix semantic verification (are -edy/-eey really prepared/fresh? Hypothesis.)
• Carbon dating: vellum is 1404–1438 AD, but the RECIPE could be much older
• The language was designed — by whom, and for what community?
Everyone asked what language it was written in.
Nobody asked what the book was for.
The book told us.
We mapped distinctive vocabulary per manuscript folio using TF-IDF. Words that appear on only ONE folio are likely the name of the plant drawn on that page.
• Folio 46r — Shee, ShecKhy, fchdy. ROOT-class ingredient with METHOD-K processing. Most distinctive page in the manuscript.
• Folio 34v — chdaiin, olchdaiin. HERB-class, “d” core, full-measure. Unique to this page.
• Folio 41r — cheked. HERB with unusual double-core “ek+ed.” Only here.
• Folio 1r — okan. SUBSTANCE-A class. The first plant in the book.
• Folio 51r — cKheody. METHOD-K processing that “becomes” something.
Each of these folios has a specific plant illustration. Cross-referencing the distinctive word with the drawing gives you the ingredient name. That’s the Rosetta Stone approach — illustration + unique word = identified ingredient. Not done yet. But the signatures are mapped.
Lines that start with TAKE (daiin) and end with a terminator (-am, -dy). These are complete instructions:
daiin okor or ol cKhol chor chom dam or Sho chol dam
TAKE SUB_A and of METHOD_K HERB HERB END and ROOT HERB END
"Take substance-A and, with method-K, herb, herb. Done.
And root, herb. Done."
daiin SheoIKhy ykey Sheky qokal qokeey okain okar ol dam
TAKE ROOT fresh ROOT PROC1 PROC1 SUB_A SUB_A of END
"Take root fresh, root, process, process,
substance-A, substance-A, of. Done."
Two sub-recipes separated by END markers. Ingredient pairs that repeat across recipes (Shed+ched appears 12 times together) confirm real combinations — like flour and water always showing up together because that’s how you make dough.
It’s an old book about medicine. Written in a notation system designed for precision by people who needed exact recipes and couldn’t afford ambiguity. The language wasn’t hidden to be mysterious. It was built to be clear — to anyone who knew the system.
37,919 words. 600 years. 87.8% structure decoded.
Not the words yet — the grammar.
The grammar says: recipe book. Medicine. Plants and roots.
An old book about how to help people.
Good will applied forward.
K = lag-1 autocorrelation on word-length signal. Measures how strongly adjacent words predict each other’s structure. High K = formulaic (recipes). Low K = variable (descriptions). Three regimes identified: 0.01–0.05 (herbal), 0.10–0.15 (catalog), 0.30–0.48 (pharmaceutical).
• 37,919 words, 8,152 unique. Vocabulary density: 0.215 (English: 0.05–0.15)
• Word-level Shannon entropy: 10.47 bits/word (English: 9–11)
• Character-level entropy: 3.95 bits/char (English: 4.0–4.5)
• Second-order conditional entropy (h2): ~2.0 (natural languages: 3.0–4.0) — anomalously low because characters are positionally constrained
• K (lag-1 autocorrelation): 0.1675 (English: 0.05–0.15, random: ~0.00)
• Zipf-compliant with heavy tail (technical vocabulary, not general prose)
• Source: Takahashi EVA transcription via github.com/alephmembeth/voynich
| Character | Total | Start % | Middle % | End % | Constraint |
|---|---|---|---|---|---|
| q | 5,422 | 99.4% | 0.6% | 0.0% | START ONLY |
| S | 4,501 | 71.9% | 28.1% | 0.0% | START-HEAVY |
| e | 20,067 | 0.7% | 98.9% | 0.4% | MIDDLE ONLY |
| h | 17,856 | 0.0% | 99.6% | 0.4% | MIDDLE ONLY |
| i | 11,660 | 0.1% | 99.9% | 0.1% | MIDDLE ONLY |
| k | 9,983 | 11.5% | 87.8% | 0.6% | MIDDLE ONLY |
| n | 6,137 | 0.0% | 1.3% | 98.7% | END ONLY |
| y | 17,504 | 10.0% | 2.8% | 87.2% | END ONLY |
| m | 1,105 | 0.2% | 4.8% | 95.0% | END ONLY |
| r | 7,358 | 5.6% | 18.5% | 76.0% | END-HEAVY |
Word = [prefix] + root + [suffix]. Closed suffix system. Same suffixes appear on all roots in same proportions.
ch- (5,684) · Sh- (2,945) · qok- (2,927) · ok- (2,365) · ot- (2,290) · qo- (2,087) · ol- (975) · cTh- (368) · dch- (248) · ych- (232) · tch- (227) · cKh- (151) · cPh- (113) · cFh- (26)
-dy (2,750) · -aiin (2,564) · -edy (2,030) · -ol (2,019) · -ar (1,707) · -or (1,300) · -al (1,238) · -ain (1,024) · -ody (861) · -eey (843) · -chy (627) · -am (530) · -eedy (521)
qokeey (308) · qokeedy (305) · qokain (279) · qokedy (272) · qokaiin (262) · qokal (191) · qokar (152) · qokol (104) · qokchy (69) · qokor (36) · qokam (25)
Same suffix distribution as ch-, ok-, ot-, Sh-. All roots take the same endings. Closed grammar.
K measured in 500-word windows across the full manuscript:
• Words 0–3000 (herbal): K = 0.03–0.05. Low coupling. Descriptive text.
• Words 3500–7000: K = 0.10–0.17. Rising structure. Transitional.
• Words 8500–9000: K = 0.33. Recipe section begins.
• Words 10000–10500: K = 0.48. Strongest coupling. Most formulaic.
• Words 11000–11500: K = 0.37.
• Words 26000–30000: K = 0.18–0.23. Pharmaceutical section. Structured but variable.
• Vocabulary overlap between adjacent sections drops to 12.8% at major boundaries.
ychor — 100% line-start (16/16) · ycheol — 93% · dShedy — 89% · ycheo — 87% · pol — 83% · tchedy — 70% · sor — 65%
ram — 100% line-end (14/14) · ary — 88% · am — 74% · qokam — 84% · okam — 73% · dam — 64% · dan — 70%
• Function words identified: 17,380 (45.8%)
• Structure decoded: 15,929 (42.0%)
• Unknown: 4,610 (12.2%)
• Total structural coverage: 87.8%
• Bowern & Lindemann (2021), “The Linguistics of the Voynich Manuscript,” Annual Review of Linguistics
• Montemurro & Zanette (2013), “Keywords and Co-Occurrence Patterns in the Voynich Manuscript,” PLoS ONE
• Stolfi (2000), Word structure grammar — crust/mantle/core model
• Ando et al. (2014), Psychotic traits in comedians, BJPsych (personality constraint model)
• Takahashi EVA transcription via github.com/alephmembeth/voynich
• Carbon dating: University of Arizona AMS, 1404–1438 AD (95% CI)
• Beinecke Rare Book & Manuscript Library, Yale University, MS 408
• GUMP tools: Sensor (K/R/E/T), Oracle (spectral), Knowledge (graph), Sfumato (entropy)