The Voynich manuscript – modern research

The previous post followed the attempts to read the Voynich manuscript, from its seventeenth-century owners to the claimants of the last few years. By the late 1970s the professionals had given up on reading it. They turned instead to measuring it: if the text could not be read, perhaps its statistics could at least say what kind of text it was. This post follows that turn, which has been the main line of serious work for fifty years.

It is a story about methods. For each one I describe what is measured, how, and what it showed when it was first applied. What the results add up to, and which explanations of the manuscript survive them, is the subject of the next post.

It goes in this order:

1. The map again, opened on the last twenty-five years.
2. From photographs to data: the transcriptions everything is computed on, with a list of those in use today and links to them.
3. Letters: entropy, and why the text is so predictable.
4. Words: frequencies, lengths and word grammars.
5. Position: lines, paragraphs and the edges of words.
6. Meaning without reading: keywords, networks and topics.
7. Machines that write Voynichese: the generators.
8. Machine learning, and the language-model wave.
9. The benchmark turn of the last two years.
10. What the graph shows about the field as a whole.
11. What is beyond doubt: the results, sorted by how firmly they stand. This is the section to read if you read only one.
12. What remains open: the questions measurement has not answered.

The map again

The citation graph from the previous post, which I compiled for this series, opened this time on the years from 1999. The old works are still there, off to the left; drag to reach them.

This figure is drawn in the browser and needs JavaScript.

Ctrl/⌘ + scroll to zoom · drag to pan · double-click to reset

Most of the graph lies in this window. The statistics lane is by far the busiest, and most of its arcs run to two places: back to Currier and the other early observers, and across to the generators, the programs and procedures that try to produce Voynich-like text. That conversation between measurement and imitation is the modern field, and this post is mostly about it.

The graph sorts the literature into five areas of research, explained in full in the previous post and summarised in the key under the graph. From the top:

• Object: the physical book. Dating, materials, imaging, the scribes, the order of the leaves and the owners. Its evidence is independent of the text.
• Surveys: surveys and transcriptions. What has been tried, and the text in a form that can be computed on.
• Statistics: statistics and linguistics. Measuring the text as it stands and comparing it with languages.
• Generators: ciphers, grilles and copying rules built to produce Voynich-like text, and then compared with it.
• Claims: works that say they have read the text.

This post is mostly about the middle three, and about the traffic between them: statistics test the claims, generators answer the statistics, and the statistics in turn test the generators.

From photographs to data

Nothing can be computed on a manuscript. It has to be turned into text first, and every choice made in doing so ends up inside every number computed from it.

The first machine-readable transcription was made by William Friedman’s study group in 1944–46, in a code designed for punched cards. Others followed in the codebreaking world, Currier’s among them. Each used its own alphabet, and each transcriber decided differently where the uncertain cases fell: whether two strokes were one letter or two, whether a narrow gap was a space. In the early 1990s members of a new Voynich mailing list typed in and extended Currier’s transcription, and Jim Reeds recovered Friedman’s, which had been lost, and published it in 1995. From then on anyone could compute on the text.¹

The decisive step was EVA, the European (later Extensible) Voynich Alphabet, designed around 1998 by Gabriel Landini and René Zandbergen, and the complete transcription in it by Takeshi Takahashi in 1999.² EVA was built to avoid deciding things in advance. It records strokes, not letters: the “bench” is written as two letters, ch, and a run of short strokes as iin, so that a researcher who thinks these are single letters can merge them, and one who does not can leave them apart. Since then nearly every statistic on the manuscript has been computed in EVA. Zandbergen’s later transcription, made with Landini and recorded in a file format that marks every uncertain reading and every uncertain space, is now the reference.

The transcriptions in use today

Five independent transcriptions are in circulation, in eight files. René Zandbergen has converted all of them to one format, IVTFF, in which every line carries its folio and its position on the page, every page its Currier language and its scribe, and every uncertain reading is marked. All eight files can be downloaded from voynich.nu, and the voynichese.com browser shows the Takahashi text highlighted on the page images.

The choices still matter, and more than is often admitted. Whether ch counts as one letter or two changes the letter statistics substantially. Whether an uncertain space is a space changes the number of distinct words. An audit published this year found that the choice of glyph representation alone moves the most-quoted letter statistic, described next, by more than a third of a bit.³ A number reported without its transcription and its rules for uncertain cases cannot be compared with any other.

Letters

Entropy

The first property the computer showed was that the text is unusually predictable at the level of letters.

The measure is the conditional entropy of a letter given the one before it. If $X_{n-1}$ and $X_n$ are consecutive letters, it is

$$\begin{array}{}\text{(}\text{1}\text{)}& h_2=H(X_n\mid X_{n-1})=-\sum _{a,b}p(a,b){\log }_{2}p(b\mid a),\end{array}$$

measured in bits. It answers the question: once you know a letter, how uncertain are you, on average, about the next one? A text in which any letter could follow any other has a high $h_2$. A text in which each letter strongly constrains the next has a low one.

In 1976 the Yale physicist William Ralph Bennett, one of the inventors of the helium–neon laser, used the manuscript as an exercise in a textbook on problem solving with computers, and found its $h_2$ remarkably low.⁴ Measured carefully, it comes out at about 2.1 to 2.5 bits, depending on how the strokes are grouped into letters, against 3.4 for English.⁵ Bennett noticed that the manuscript sat close to a Hawaiian text he had measured, and Hawaiian, with its small alphabet and syllables that never end in a consonant or put two consonants together, is about as predictable as written languages get. That looked encouraging. Dennis Stallings later pointed out that Bennett’s Hawaiian sample had been written in a simplified spelling that left out the glottal stop and the length of vowels, which made it look more predictable than Hawaiian is.

In Lindemann and Bowern’s corpus of 311 Wikipedia languages, Hawaiian is still the lowest. The values run from 6.14 bits (for Chinese characters; alphabets run from about 2.8 to 3.8) down to 2.77–2.84 for the four most predictable: Hawaiian, Venda, Tswana and Min Dong, written in Latin letters, all but Min Dong with simple syllables, and Venda and Tswana tonal languages whose spelling leaves the tones out. The manuscript sits below all of them, but much nearer to them than to English.

This figure is drawn in the browser and needs JavaScript.

Ctrl/⌘ + scroll to zoom · drag to pan · double-click to reset

Low $h_2$ became the central anomaly of the field. Whatever the manuscript is, an explanation has to say why its letters are so tightly constrained. The definitive study came in 2020, when Luke Lindemann and Claire Bowern of Yale measured $\text{(1)}$ on 311 Wikipedia samples in many languages and on 18 historical texts, and on the manuscript in Takahashi’s transcription under three ways of grouping its strokes into letters.⁵ The low value survived all of them. It is not an artefact of EVA, it holds in both of Currier’s languages and in every scribe’s hand, and none of their groupings brings it up to the range of the 311 languages they measured, though the margin to the nearest language is smaller than the effect that the way of counting glyphs can have, and the languages were counted letter by letter. Abbreviation, which one might expect to explain it, does the opposite: abbreviated Latin is less predictable than Latin written out. Their reading was that the placement of letters inside a word is highly constrained, as if the script recorded fewer distinctions than the language underneath it makes.

Vowels, and the abjad idea

If the letters are so constrained, perhaps some of them are vowels and some consonants, alternating as they do in speech. In 1991 Jacques Guy applied to the manuscript an algorithm devised by the Soviet linguist Boris Sukhotin, which separates the letters of an unknown alphabetic text into vowels and consonants from nothing but which letters tend to stand next to each other. It found four signs that behaved like vowels, but the separation was weak, and many words contain none of them.⁶

In 2011 Sravana Reddy and Kevin Knight, two computational linguists, fitted a hidden Markov model with two states to the letters.⁷ Run on English, such a model splits the letters into vowels and consonants. Run on the manuscript, it split off the last letter of each word from all the others. Arabic written without its vowel marks, and English with its vowels removed, give the same split, and from this they proposed that the script is an abjad, a script that writes consonants and leaves vowels out, as Hebrew and Arabic do. Lindemann and Bowern’s measurements later counted against the idea: scripts written without vowels have more uncertain letters than alphabets, not fewer, so an abjad moves the entropy the wrong way.

Words

Frequencies and Zipf’s law

In almost every text in any language, the second commonest word occurs about half as often as the commonest, the third about a third as often, and so on. This is Zipf’s law: if $f(r)$ is the frequency of the word of rank $r$, then

$$f(r)\propto r^{-\alpha },\alpha \approx 1.$$

The manuscript obeys it, and for a while that was taken as evidence that it is a language. It is not much evidence. George Miller pointed out in 1957, and Wentian Li showed in detail in 1992, that text typed by a monkey at random, with a space bar among the keys, obeys something close to Zipf’s law too.⁸ Most processes that make words out of letters come close.

Here is what that looks like. The figure shows three texts of the same length, about 35,000 words each: the manuscript, a Latin botanical text, and random letters with a space thrown in about one time in five or six.

This figure is drawn in the browser and needs JavaScript.

Ctrl/⌘ + scroll to zoom · drag to pan · double-click to reset

The random text falls in steps, one for each word length: all one-letter words are about equally common, then all two-letter words, and so on. A line through the corners of the steps has a slope close to −1, which is Li’s point. Whatever the figure shows, it does not show that the manuscript is a language.

A stronger argument came in 2001 from Landini, who removed every space from the transcription and looked at the correlations in the resulting stream of letters with the tools of spectral analysis.⁹ A random stream has none. Language has a characteristic structure of correlations over short and longer distances, and so did the manuscript. The analysis also recovered a typical length of the units in the stream without trusting the spaces at all, and it agreed with the words. It was the first strong statistical argument that the text is not a random sequence of signs.

Word lengths and word grammars

Jorge Stolfi, a computer scientist in Brazil, published a series of studies on the web between 1997 and 2000 that are still among the most cited in the field.¹⁰ He showed two things. First, the lengths of the manuscript’s words follow a nearly symmetric, binomial distribution. Latin and the other European languages have a long tail of long words, and the manuscript has almost none. Stolfi later found binomial word lengths in Chinese written in pinyin, in Tibetan and in Vietnamese, scripts that space syllables rather than words, so the shape rules out a European alphabet rather than language as such, and suggests the manuscript’s “words” may be closer in size to syllables. Second, he showed that almost every word fits a single scheme, which he described as a core, wrapped in a mantle, wrapped in a crust, each layer drawn from its own small set of letters. It is Tiltman’s observation, published in 1967, that words are made of beginnings and endings, turned into a grammar that can be written down and tested.

This figure is drawn in the browser and needs JavaScript.

Ctrl/⌘ + scroll to zoom · drag to pan · double-click to reset

Later grammars refined it. The most influential recent one, by Massimiliano Zattera in 2022, describes a word as a sequence of slots, each of which may be empty or filled from a short list, in a fixed order.¹¹ A grammar like that accounts for most of the words in the manuscript. That is unlike the words of Latin or the other European languages the manuscript is usually compared with. It is more like the syllables of some Asian languages, or like languages whose words are built from fixed slots, which is one reason the slot grammar has not settled what the words are.

Position

Currier had said in 1976 that the line is a functional unit. The computer made it possible to check, and the checks have only made the effect clearer.

The first and last words of lines behave differently from words in the middle, in the letters they contain and in their length. Certain letters prefer the start of a line, others its end. The tall “gallows” letters p and f occur mostly in the first line of a paragraph. Elmar Vogt re-measured the line effect in 2012, and in 2022 Patrick Feaster gave the first systematic treatment of how letter frequencies change from left to right along a line and from top to bottom down a paragraph.¹²

The most consequential measurement of position, though, was about the edges of words. Currier had noticed in 1976 that some word endings are followed by particular beginnings. Sravana Reddy and Kevin Knight measured it in 2011, on a measure of predictability that did not separate it clearly from English: one letter of the previous word helped predict the next word’s first letter by about as much in the manuscript as in English.¹³ In 2019 Emma May Smith and Marco Ponzi gave the first systematic treatment: the last letter of one word and the first letter of the next are strongly dependent, with certain endings followed by certain beginnings far more often than chance would allow.¹⁴ With shuffled text as the baseline, the latest study puts the effect well above the European prose it tested (Rozanova and Temerev, below). This dependency at the word edge has become one of the two sharpest tests that any model of the text has to pass; the other, the number of words that occur only once, comes up below.

Meaning without reading

Can statistics show that a text means something without reading it? Several methods have tried.

In 2013 Marcelo Montemurro and Damián Zanette used information theory to find the manuscript’s keywords: the words whose occurrences cluster in particular parts of the text rather than being spread evenly through it.¹⁵ In a meaningful text the words that carry the subject behave that way: a book about plants mentions “leaf” in the chapters about leaves. They found that the manuscript’s keywords clustered by section, and that the sections defined by the drawings were also sections in the statistics of the words. Their paper is the most cited piece of Voynich research in the scientific literature, according to the OpenAlex index.

In the same year Diego Amancio and his colleagues turned texts into networks, with words as nodes linked to their neighbours, and calibrated measures of network shape and of how words recur on the New Testament in fifteen languages.¹⁶ The manuscript came out “mostly compatible with natural languages and incompatible with random texts”.

In 2021 Rachel Sterneck, Annie Polish and Claire Bowern applied topic modelling, a method that finds groups of words that tend to occur together, and found that the pages cluster by scribe and by subject together.¹⁷ In 2026 Colin Layfield and Lisa Fagin Davis used a related method, latent semantic analysis, on short chunks of text, and found that neighbouring chunks are more alike than distant ones, as in a meaningful text, with drops at page and section breaks.¹⁸ They went further, and argued from the same method that each bifolium was once a self-contained unit of text. That conclusion is disputed, and the book’s original order is a question in its own right.

The fullest linguistic case is Claire Bowern and Luke Lindemann’s survey of 2021.⁵ On measures of vocabulary they found the manuscript inside the range of natural languages: its variety of words in moving windows sits in the middle of their language families, closest to Iranian, Germanic and Romance, and words repeated twice in a row are no more frequent than in some natural texts. They read Currier’s line effects as language too, suggesting that a line may roughly correspond to a sentence. Their conclusion was that the letters do not stand one for one for sounds, but that the words and lines behave like “regular natural language”.

All these methods share a limit, and it is the subject of the next section. They show that the text is not random, and that its words are distributed in ways meaningful text also shows. Whether a text without meaning, produced in the right way, would show the same is a separate question.

Machines that write Voynichese

The grille

In 2004 Gordon Rugg, of Keele University, put the hoax hypothesis on a new footing. Earlier versions had asked who might have forged the book and why. Rugg asked how.¹⁹ He showed that a variant of a sixteenth-century device, the grille Girolamo Cardano described in 1550, could do it: a card with holes cut in it, slid over a table of syllables, picks out beginnings, middles and endings, and a scribe copies down whatever shows through. Text made that way has words with regular structure, repeats, and a convincing look, and it can be produced quickly by hand. Rugg did not claim the manuscript had been made like that. He claimed that it could have been, and that the burden lay on anyone who said it meant something. He wrote before the parchment was dated, though. The radiocarbon result puts it more than a century before Cardano, so the grille is a period procedure only if the parchment was old when it was written on, or if something like it was in use earlier, for which there is no evidence. Rugg returned to the question with Gavin Taylor in 2016, showing that other low-level statistics of the text can also be produced by hand.²⁰

The paper started what is now the second busiest lane of the graph. Its real contribution was methodological. A hypothesis about meaningless text had become a procedure, and a procedure can be run, and its output measured against the manuscript. By that test, the grille itself does badly: in one recent comparison, a preprint, it passed one test of eight, and that one is passed by seventeen of the eighteen models tested.²¹ But it changed the question.

Self-citation

Andreas Schinner had argued in 2007, from random-walk and repetition statistics of the text, that it was produced by a stochastic process rather than written as language.²² The most successful generator comes from Torsten Timm, who proposed in 2014 that each word in the manuscript was made by copying a word already written nearby and changing it slightly.²³ The scribe looks back up the page, takes a word, alters a letter or two, and writes it down. Over thousands of words this produces a text full of near-repeats, with a small set of building blocks and a vocabulary that drifts slowly through the book. With Andreas Schinner he published the procedure in 2020 as a program anyone can run.²⁴ It matches the low entropy, the regular words and the weak dependence between consecutive words as it is usually measured. It falls short on how many words occur only once, and on the dependency at the word edges. In 2021 Timm and Schinner replied to Bowern and Lindemann’s case for language in a short review.²⁵

Gibberish by hand

In 2022 Gaskell and Bowern ran an experiment.²⁶ They asked volunteers to write pages of meaningless text by hand, and compared the results with the manuscript and with meaningful texts on forty-two statistical measures. By a loose test of similarity, at least 5 per cent overlap between the distributions, the manuscript resembled the gibberish on 41 of them and meaningful text on 38; by the same test, the gibberish resembled meaningful text on all 42. A classifier trained on the samples put the manuscript nearer the gibberish, with low confidence, and the samples were short. The conclusion was not that the manuscript is gibberish. It was that most of the statistics the field had relied on cannot tell meaningful text from meaningless, so they cannot settle the question either way. A companion paper by Bowern and Gaskell, “Enciphered after all?”, found the word-level statistics compatible with some kinds of cipher too.²⁶

A cipher that works

The most important recent generator is a cipher. In 2025 Michael Greshko published the Naibbe cipher, which turns Latin or Italian into Voynich-like text by hand, using nothing but playing cards, dice and a set of tables.²⁷ The plaintext’s spaces are removed, and a throw of two dice splits its letters into single letters and pairs. A card drawn from a shuffled deck then picks one of six tables. A single letter becomes one whole Voynich-like word, and a pair becomes a prefix for its first letter followed by a suffix for its second. The tables are built so that every such word obeys the manuscript’s slot grammar. It was the first substitution cipher shown to reproduce many of the manuscript’s word-level properties at once, with means available in the fifteenth century. Greshko published his code and tables; Rozanova and Temerev, below, reimplemented it independently, and recent studies use it as a control.

It does not reproduce everything. Like self-citation, it falls short on the number of words that occur only once and on the coupling at word edges, and it has no mechanism for the line. Across its word breaks Rozanova and Temerev measured almost no coupling (0.006 bits against the manuscript’s 0.2): whatever links one plaintext letter to the next does not survive into the glyphs at the edges of the cipher’s words. These properties, together with the line structure, are what the best imitations still fail to reproduce.

Machine learning

Machine learning has been applied to the manuscript in two quite different ways.

The first is as a tool for measurement. Reddy and Knight’s hidden Markov model was an early case, and so was Bradley Hauer and Grzegorz Kondrak’s method of 2016 for identifying the language behind an unknown script, which was sound even though its reading of the manuscript as Hebrew was not.²⁸ More recent work trains language models from scratch on the transcription and studies what they learn: which earlier words a model attends to when it predicts the next, and whether that depends on position in the line. A preprint published this year set eighteen model variants, from nine families, against the manuscript on eight measures at once, from simple random samplers through Markov chains to a neural network fine-tuned on the text. None reproduced all eight. The best, passing seven as his released code counts them (six under the two-sided criterion his paper states), was a simple sampler of word pairs with no content at all, given the book’s sections in advance.²¹ This is useful work. It treats the models as instruments, and its claims can be checked.

The second is as an oracle. Since 2025 hundreds of claimed decipherments have been deposited in open repositories, many of them produced in conversation with large language models. A typical one builds a dictionary, adjusts it passage by passage, asks the model whether the resulting translation is coherent, and reports rising “coverage” as progress. This repeats the oldest failure in the history of the manuscript, the key adjusted to fit the text, with a new and very fluent accomplice. A language model asked whether a translation makes sense will generally say yes, because producing plausible prose from weak constraints is what it is built to do. Coverage of a dictionary enlarged to maximise coverage measures effort, not correctness.

The benchmark turn

In the last two years the field has changed its habits, and the change is visible in the graph as a crowd of new works in 2025 and 2026.

The change is about how claims are tested. In the best recent papers, a statistic of the manuscript is no longer compared with a few European languages, or with a random shuffle. It is compared with matched controls at the same length: prose in many languages, ciphers, gibberish, and the generators described above, rebuilt by the authors of the new studies rather than taken on trust. Results are reported for each of Currier’s languages separately, and resampled by quire, leaving whole gatherings out in turn, so that a few unusual pages cannot drive them. Some studies are preregistered: the test is fixed and announced before the data are looked at. Negative results are published.

Three examples show the style. Christophe Parisel published in 2026, in preprints, the first benchmarks set separately for each of Currier’s languages and evaluated jointly, and showed that neither the grille nor the slot grammar reproduces them together.²⁹ Rod Kinnison found in the same year that two properties of the letters within words, which none of his four language controls (English, Latin, Finnish and the constructed Esperanto) matches, are matched by the Naibbe cipher applied to his Latin source without any tuning: the first peer-reviewed test of that kind of cipher, and a result consistent with it. He did not test a generator of meaningless text.³⁰ And in August 2026 Liudmila Rozanova and Alexander Temerev, at IIASA in Austria, published as a preprint what is probably the most careful paper of the period, testing the three assumptions built into the field’s vocabulary: that a glyph is a letter, that a group of glyphs between spaces is a word, and that a space is a word space.³¹ Against matched controls, all three failed. The glyphs are too regular to be the letters of any plaintext they tested. The groups between spaces predict one another far less than the words of any control text. And the spaces the transcribers marked as uncertain behave like breaks inside a word, not between words. They did not stop using the words “glyph”, “word” and “space”, and neither does this series: the terms are labels for what is on the page, and what they stand for has to be shown, not assumed. They also reimplemented the two leading generators and named the two properties that still separate the manuscript from both: how many of its words occur only once, and how strongly the end of one word constrains the start of the next.

That is where the computational study of the manuscript stands. Fifty years of measurement have ruled out a great deal: random text and simple substitution of a European language, and they have made very unlikely an ordinary abjad and, in two preprint implementations, Rugg’s grille. They have also shown that the statistics the field once leaned on cannot decide between meaning and its absence. What is left is a short list of properties that no model reproduces, and a much better idea of how to test the next one.

What the graph shows

The graph also says something about the field as a whole, and some of it is surprising. The counts below are my own, made on the graph compiled for this series. One limit first: 24 of the 69 works, most of them early, have no reference list that could be read, so they appear only as cited works, and counts of outgoing citations favour recent papers for that reason.

The field stands on a few works. The most cited works within the graph are Currier’s papers of 1976, cited by 21 of the others; then Rugg’s grille, Timm and Schinner’s generator and Montemurro and Zanette’s keywords, with 19 each; then Landini, Bowern and Lindemann’s review, and Reddy and Knight. Two of the top four are generators. Every statistical study has to answer the hypothesis that the text means nothing, so it cites it.

Measurement and imitation are in constant conversation. Of the 329 citations, 95 run within the statistics lane, 49 from generators to statistics, and 44 from statistics to generators. No other pair of lanes comes close. Every measurement is tested against a mechanism, and every mechanism against the measurements.

The physical evidence is almost never cited. The radiocarbon dating and the materials analysis, the two most decisive facts about the object, receive no citations at all in any reference list that could be read. Statistical papers treat the date as common knowledge. It is a symptom of a real separation: the new codicological work, which questions the order of the pages, has not yet reached the statistical studies whose measures of adjacency it bears on.

Decipherment claims are cited as warnings. The claims lane receives 35 citations, 22 of them from statistical papers, nearly all in introductions listing failed solutions. Newbold and Manly are still its most cited works, a century on.

The recent work reaches widest. The papers of the 2020s cite across every lane and every era. Rozanova and Temerev’s cites 27 works in the graph, more than any other. Across the whole graph the median citation reaches back eleven years.

What is beyond doubt

Fifty years of measurement have produced results of very different firmness, and the field has learned to say which is which. Here are the results that matter, sorted by how firmly they stand. The levels used here are these: established means physical evidence, or measured repeatedly by independent groups on different transcriptions, and not seriously disputed; strong means peer-reviewed or independently reproduced with controls; supported means one careful study with controls, or several weaker ones agreeing, not yet replicated; contested means serious work on both sides. Only published work, preprints included, is counted here; a result from a single preprint is at most supported.

Established.

• No decipherment has survived independent testing, in more than a century of attempts.
• The book is a genuine medieval object: its parchment dates to 1404–1438, its materials are of the period, and it was in Prague by 1639. That it was written soon after the parchment was made is a strong inference, not a measurement.
• The letters are unusually predictable, far more so than in English or any European prose measured (3.4 bits and more), whatever the transcription and recipe: about 2.1 to 2.7 bits (Bennett, 1976; Lindemann and Bowern, 2020; Rozanova and Temerev, 2026).
• The text is not random, by every measure applied since 2001 (Landini, 2001, and all later work).
• It is not a simple substitution of Latin, Italian or English. No one-to-one assignment of letters can turn their statistics into the manuscript’s.
• Words follow a rigid internal grammar, built from parts in a fixed order (Tiltman, Stolfi, Zattera).
• Word lengths are narrow and near-binomial, with almost no long words, when uncertain spaces are read as spaces (Stolfi, 2000); a preregistered audit found the claim held only conditionally, failing mainly when they are not (Lux, 2026).
• Position governs the choice of glyph: in the word, in the line and in the paragraph (Currier, 1976; Vogt, 2012; Feaster, 2022).
• The end of a word constrains the start of the next (noticed by Currier, 1976; measured by Reddy and Knight, 2011; studied by Smith and Ponzi, 2019).
• Zipf’s law holds, and it is weak evidence, because random text comes close to it too (Landini, 2001; Miller, 1957; Li, 1992).

Strong.

• Two regimes, Currier’s A and B, recovered without labels by modern methods (Currier, 1976; Parisel, 2026). Whether they are two categories or the ends of a spectrum is argued over, and one study finds no evidence of two languages at all (Ponnaluri, 2025).³²
• Vocabulary follows subject: keywords cluster by section, and pages cluster by scribe and topic (Montemurro and Zanette, 2013; Sterneck, Polish and Bowern, 2021). It is the main statistical argument for meaning. Sections largely coincide with scribes and with Currier’s A and B, so the clustering is not yet separated from them.
• No regrouping of the glyphs gives the letter statistics of any tested language (Lindemann and Bowern, 2020), though the margin depends on how the glyphs are counted.

Supported.

• About 70 per cent of word types occur only once in the paragraph text, more than either of the two generators tested on it (Rozanova and Temerev, 2026).
• The coupling at the word edge is about 0.2 bits after correction for chance, above eight European prose texts and above both generators tested (Rozanova and Temerev, 2026); one self-published cipher claims to reach it.
• Rozanova and Temerev argue that a glyph is not a letter of any plaintext they tested, and that the units of the text are groups of two or three glyphs (2026).
• Neighbouring words tend to be alike in length, measured over the text as a whole, across line breaks; a habit the authors found common in handwritten gibberish and rare in language (Gaskell and Bowern, 2022).
• Word order is weak. One word predicts the next much less than in the control texts of the studies that measured it (Reddy and Knight, 2011; Rozanova and Temerev, 2026), on a small, noisy estimate that moves with the number of word types kept and is about zero within language A or B alone.
• The best mechanisms fail in two known places. The Naibbe cipher and the self-citation generator reproduce most of the statistics, but neither reaches the 70 per cent of single-use words or the coupling at word edges (Rozanova and Temerev, 2026).

Contested.

• Five scribes (Fagin Davis, 2020), or one evolving hand with labels and two regimes (Timm, 2026).
• The book as a set of independent bifolia, later bound out of order (Layfield and Fagin Davis, 2026), against critics who accept the similarity within bifolia but not the conclusion.
• Language or meaningless text. Bowern and Lindemann, and Montemurro and Zanette, on one side; Timm and Schinner, who replied to Bowern and Lindemann in 2021, on the other. Gaskell and Bowern’s experiment shows that the low-level statistics cannot decide between them.

What remains open

And here is what fifty years of measurement have not answered, roughly from the questions that can be settled with a computer to those that need institutions or new ideas.

1. Meaning or not? Four families of explanation survive: a cipher more elaborate than any known from the period, a language written in a script that records fewer distinctions than the language makes, an unrecorded or invented language, and meaningless text from a procedure. A heavier shorthand and a notation remain possible but have hardly been tested. The measurements have not decided between them.
2. What produces the rich vocabulary and the coupling at word edges together? Neither of the two leading mechanisms reproduces both; a self-published cipher claims the coupling.
3. How does Currier’s A relate to B? Same script, different favourite words and letters, and nobody has shown how one maps onto the other.
4. What is the line for? Every measurement since Currier has confirmed that the line is a unit. None has explained why.
5. What are the recurring parts of words? Linguistic morphemes, units of a cipher, habits of the pen, or something else.
6. Could a cipher worked by hand carry structure over long distances? The best ciphers so far choose their variants at random, letter by letter.
7. Would a heavier shorthand than medieval practice fit? Ordinary Latin abbreviation raises letter entropy instead of lowering it, its other effects have hardly been measured, and nothing heavier has been built and tested.
8. What was the original order of the leaves, and were bifolia read as units?
9. What language are the month names? A small, old, still unanswered question.
10. What would imaging the rest of the book show? Only ten pages have been photographed multispectrally, and those ten revealed the table in the margin of folio 1r. That needs the library, not cleverness.

What comes next

The next post gathers the measurements described here into a harness: the properties any explanation of the text has to reproduce, and the rules, starting from the habits of “The benchmark turn”, for measuring them fairly. Each concept in the search space laid out in the previous post can then be held to it.

Footnotes

1. J. Reeds, “William F. Friedman’s Transcription of the Voynich Manuscript”, Cryptologia 19/1 (1995), 1–23. On the mailing list’s work on Currier’s transcription, see voynich.nu. ↩

2. G. Landini and R. Zandbergen, EVA (1998–2000), and T. Takahashi’s complete transcription (1999); see voynich.nu for both and for the Zandbergen–Landini transcription that replaced them as the reference. ↩

3. V. Lux, “How Stable Are Quantitative Claims about Voynichese?” (Zenodo, September 2026), doi:10.5281/zenodo.22850758; preregistered, self-published and not peer reviewed. ↩

4. W. R. Bennett Jr., Scientific and Engineering Problem-Solving with the Computer (Englewood Cliffs: Prentice-Hall, 1976). ↩

5. L. Lindemann and C. Bowern, “Character Entropy in Modern and Historical Texts: Comparison Metrics for an Undeciphered Manuscript”, arXiv:2010.14697 (2020); C. Bowern and L. Lindemann, “The Linguistics of the Voynich Manuscript”, Annual Review of Linguistics 7 (2021), 285–308. ↩ ↩² ↩³

6. J. B. M. Guy, “Statistical Properties of Two Folios of the Voynich Manuscript” and “Vowel Identification: An Old (but Good) Algorithm”, Cryptologia 15/3 (1991), 207–218 and 258–262. ↩

7. S. Reddy and K. Knight, “What We Know About the Voynich Manuscript”, Proceedings of the 5th ACL-HLT Workshop on Language Technology for Cultural Heritage, Social Sciences, and Humanities (Portland, 2011), 78–86. ↩

8. G. A. Miller, “Some Effects of Intermittent Silence”, American Journal of Psychology 70 (1957), 311–314; W. Li, “Random Texts Exhibit Zipf’s-Law-Like Word Frequency Distribution”, IEEE Transactions on Information Theory 38/6 (1992), 1842–1845. ↩

9. G. Landini, “Evidence of Linguistic Structure in the Voynich Manuscript Using Spectral Analysis”, Cryptologia 25/4 (2001), 275–295. ↩

10. J. Stolfi, “A Grammar for Voynichese Words” and “On the VMS Word Length Distribution” (2000), and earlier notes from 1997, published on his pages at the University of Campinas. ↩

11. M. Zattera, “A New Transliteration Alphabet Brings New Evidence of Word Structure and Multiple ‘Languages’ in the Voynich Manuscript”, Proceedings of the International Conference on the Voynich Manuscript 2022, CEUR-WS 3313. ↩

12. E. Vogt, “The Line as a Functional Unit in the Voynich Manuscript: Some Statistical Observations” (2012, self-published); P. Feaster, “Rightward and Downward Grapheme Distributions in the Voynich Manuscript”, Proceedings of the International Conference on the Voynich Manuscript 2022, CEUR-WS 3313. ↩

13. S. Reddy and K. Knight, “What We Know About the Voynich Manuscript”, Proceedings of the 5th ACL-HLT Workshop on Language Technology for Cultural Heritage, Social Sciences, and Humanities (2011), 78–86, Table 5: one letter of the previous word improves the prediction of the next word’s first letter by 31.8% in Currier’s language B, 31.1% in English and 26.8% in Arabic. ↩

14. E. M. Smith and M. Ponzi, “Glyph Combinations across Word Breaks in the Voynich Manuscript”, Cryptologia 43/6 (2019), 466–485. ↩

15. M. A. Montemurro and D. H. Zanette, “Keywords and Co-Occurrence Patterns in the Voynich Manuscript: An Information-Theoretic Analysis”, PLOS ONE 8/6 (2013), e66344. ↩

16. D. R. Amancio, E. G. Altmann, D. Rybski, O. N. Oliveira Jr. and L. da F. Costa, “Probing the Statistical Properties of Unknown Texts: Application to the Voynich Manuscript”, PLOS ONE 8/7 (2013), e67310. ↩

17. R. Sterneck, A. Polish and C. Bowern, “Topic Modeling in the Voynich Manuscript”, arXiv:2107.02858 (2021). ↩

18. C. Layfield and L. Fagin Davis, “The Application of Latent Semantic Analysis to the Voynich Manuscript”, Digital Humanities Quarterly 20/1 (2026), doi:10.63744/2ezxpskcezq4, and “Singulion Structure and the Voynich Manuscript”, Digital Medievalist 19 (2026), doi:10.4000/16k0a. ↩

19. G. Rugg, “An Elegant Hoax? A Possible Solution to the Voynich Manuscript”, Cryptologia 28/1 (2004), 31–46. ↩

20. G. Rugg and G. Taylor, “Hoaxing Statistical Features of the Voynich Manuscript”, Cryptologia 41/3 (2017; online 2016), 247–268, doi:10.1080/01611194.2016.1206753. ↩

21. Y. Ou, “A Statistical Turing Test for the Voynich Manuscript”, preprint (Research Square, May 2026), doi:10.21203/rs.3.rs-9755825/v1. ↩ ↩²

22. A. Schinner, “The Voynich Manuscript: Evidence of the Hoax Hypothesis”, Cryptologia 31/2 (2007), 95–107. ↩

23. T. Timm, “How the Voynich Manuscript Was Created”, arXiv:1407.6639 (2014). ↩

24. T. Timm and A. Schinner, “A Possible Generating Algorithm of the Voynich Manuscript”, Cryptologia 44/1 (2020), 1–19. ↩

25. T. Timm and A. Schinner, review of Bowern and Lindemann, Cryptologia 45/5 (2021), 434–438, doi:10.1080/01611194.2021.1911875. ↩

26. D. Gaskell and C. Bowern, “Gibberish after All? Voynichese Is Statistically Similar to Human-Produced Samples of Meaningless Text”, and C. Bowern and D. Gaskell, “Enciphered after All?”, both in Proceedings of the International Conference on the Voynich Manuscript 2022, CEUR-WS 3313. ↩ ↩²

27. M. A. Greshko, “The Naibbe Cipher: A Substitution Cipher That Encrypts Latin and Italian as Voynich Manuscript-Like Ciphertext”, Cryptologia (online 26 November 2025), doi:10.1080/01611194.2025.2566408. ↩

28. B. Hauer and G. Kondrak, “Decoding Anagrammed Texts Written in an Unknown Language and Script”, Transactions of the Association for Computational Linguistics 4 (2016), 75–86. ↩

29. C. Parisel, “Evidence of Layered Positional and Directional Constraints in the Voynich Manuscript: Implications for Cipher-Like Structure”, arXiv:2604.19762, and “A Quantitative Confirmation of the Currier Language Distinction”, arXiv:2604.25979 (2026). ↩

30. R. Kinnison, “The Voynich Manuscript’s Positional Entropy Collapse: Comparative Evidence from Languages and Ciphers”, Cryptologia (online 7 September 2026), 1–18, doi:10.1080/01611194.2026.2697318. ↩

31. L. Rozanova and A. Temerev, “A Glyph Is Not a Letter, a Token Is Not a Word, a Space Is Not a Space”, arXiv:2608.17096 (2026). ↩

32. R. V. Ponnaluri, “The Voynich Manuscript Was Written in a Single, Natural Language”, Cryptologia 49/6 (2025; online 2024), 505–524, doi:10.1080/01611194.2024.2414128. ↩