Voynich Writing: What the Text Does Before We Can Read It

You can learn a surprising amount about writing without knowing what a single word means.

That sounds impossible only because reading is normally so fast.

You see a sentence.

Meaning arrives.

The machinery disappears.

But remove meaning and suddenly the machinery becomes visible.

Some signs appear often.

Some appear rarely.

Some combinations recur.

Some prefer the beginning of a line.

Some are unusual at the end.

Some word-like forms cluster on certain pages.

Some pages share a vocabulary profile that other pages do not.

Labels behave differently from long paragraphs.

Paragraph openings behave differently from their middles.

And suddenly the central question changes.

Not:

What does this word mean?

But:

What kind of system would make marks behave like this?

That is the doorway into the Voynich text.

Quick Read

One-sentence answer: the Voynich text is highly organised at glyph, word-like, line, paragraph, page and manuscript scales, but none of those measurable regularities currently establishes an accepted plaintext, language, cipher or semantic dictionary.

The writing system is unidentified and remains undeciphered.
Modern researchers use transcription systems to represent the visible signs consistently enough for comparison and computation.
A transcription symbol is not the original glyph and certainly not its meaning.
The number of distinct Voynich glyphs depends partly on how compound forms, ligatures and ambiguous strokes are segmented.
Spaces divide the text into useful word-like units, but we do not know whether those spaces correspond exactly to lexical words.
Some transcribed forms recur extremely often; others are strongly local.
Position matters. Line beginnings, endings, paragraph openings and other locations show different distributions.
Prescott Currier’s A/B distinction captures large-scale statistical differences in the text, but “A” and “B” are descriptive regimes, not identified languages or subjects.
Labels and running prose-like text should not automatically be treated as the same textual environment.
2013 studies in PLOS ONE found non-random, language-compatible organisation and long-range distributional structure.
Language-compatible does not mean language identified.
A 2025 Cryptologia study demonstrated that a historically plausible hand-operable cipher can reproduce several famous Voynich-like statistical properties while retaining meaningful underlying plaintext.
The five-scribe interpretation is influential but contested, so handwriting and textual variation should not be collapsed into an unquestioned author count.
The strongest surviving public conclusion is structural: the text behaves as though its producer was following constraints.

That is a much stronger statement than “random”.

It is also a much weaker statement than “decoded”.

The interesting work lives between them.

Before Translation Comes Transcription

Suppose you want to count a Voynich sign.

You cannot type the medieval mark directly into an ordinary spreadsheet and expect every researcher to agree which modern character represents it.

So researchers build transcription conventions.

One widely encountered system is EVA, the Extensible Voynich Alphabet. It assigns convenient Roman-letter labels to recurring visible forms.

This makes comparison possible.

A string such as daiin can be discussed without reproducing the original glyph shapes every time.

But this convenience creates one of the easiest misunderstandings in Voynich research.

EVA letters are labels for shapes. They are not a claim that the shapes sound like those Latin letters.

If EVA uses d, that does not mean the original sign is the consonant /d/.

If a transcribed token reads ol, that does not mean the manuscript contains the modern word “ol”.

Transcription is a map.

Do not mistake the map’s street names for the language of the territory.

Even the Number of Glyphs Is a Research Question

At first glance, counting distinct signs sounds easy.

Look.

Classify.

Count.

Then handwriting enters.

One sign can vary in shape.

Two adjacent signs can touch.

A compound-looking form may be one glyph, two glyphs or a ligature.

A damaged stroke may turn one class into another.

A flourish may be semantic, phonetic, positional or merely scribal.

Some of the tall characters conventionally called “gallows” are visually distinctive, but the nickname describes appearance rather than function.

So character inventory is partly a segmentation problem.

And segmentation matters because every frequency table depends on what we decided the units were.

This is a profound general lesson:

Before you count a thing, define what counts as one thing.

The Spaces Might Be Telling the Truth—or Only Part of It

The text contains visible spaces.

That makes the strings between spaces look like words.

Researchers therefore need some name for them.

“Word”, “token” and “word-like unit” are all used in different contexts.

The cautious form is useful because a visible space can do several things in writing systems.

separate lexical words;
separate syllabic or morphemic groups;
mark encoded units;
divide chunks created by a generative rule;
or reflect scribal conventions that do not map cleanly onto modern word boundaries.

If we decide too early that every space is a word boundary, our later statistics inherit that assumption.

That does not make token statistics useless.

It means the unit should be described honestly.

We can say:

this space-delimited form recurs 100 times

without pretending we have already proved:

this medieval word means X.

Voynich Words Look Related to One Another

One of the manuscript’s most striking textual features is how often common word-like forms resemble nearby or related forms.

A short form may gain a prefix-like element.

A central sign may change.

An ending-like sequence may vary.

Forms can look as though they belong to families.

Natural language can do this.

English has:

teach;
teacher;
teaches;
teaching;
unteachable.

But morphology is not the only process that can create families of similar strings.

A cipher can.

A copying procedure can.

A constrained generation algorithm can.

A syllabic notation system can.

So word families are important evidence about construction.

They are not self-translating.

Position Is Part of the Text

Suppose a token appears 200 times.

A frequency count tells us that it is common.

Now suppose 150 of those appearances happen near line endings.

That is a different kind of fact.

The token is not merely common.

It is position-sensitive.

The Voynich text contains many positional regularities of this general kind. Researchers have long noted that the beginnings and endings of lines are not statistically interchangeable, and paragraph openings display their own behaviour.

This is one of the reasons a “bag of words” model loses information.

If you shuffle all tokens from a folio into one bucket, you preserve frequency and destroy position.

But position may be part of the generating rule.

Where a form appears can be as important as how often it appears.

Line Beginnings Are Not Neutral Territory

In ordinary prose, line breaks can be accidental.

Change the font size and the same sentence wraps differently.

Handwritten manuscript lines are different.

The writer sees the physical edge.

The writer chooses where to begin.

The writer manages available space.

In Voynich, line-initial positions have long attracted attention because some forms and glyph classes are disproportionately associated with them.

There are many possible explanations.

linguistic syntax;
paragraph or sentence marking;
scribal abbreviation conventions;
cipher state;
decorative or emphatic forms;
line-generation procedures;
layout-driven variation.

The important discovery is not that one of these has already won.

It is that the line boundary behaves like a meaningful constraint in the production system.

Line Endings Create a Different Problem

A writer approaching the right margin has less physical room.

That can affect what gets written.

In ordinary manuscripts, scribes may abbreviate, compress, continue below or manage line length in other ways.

If Voynich line-final forms differ systematically, at least two broad explanation families become relevant.

One says the position has linguistic or encoded meaning.

The other says the physical act of filling a line changes the output.

These are not mutually exclusive.

A natural-language scribe can respond to available space.

A cipher clerk can respond to available space.

A generative procedure can encode the position directly.

So line-final behaviour is not a translation.

It is a discriminator that good theories need to explain.

Paragraph Openings Behave Like Special Places

A paragraph is more than several lines stacked together.

In many writing systems, openings carry special information.

A topic begins.

A discourse marker appears.

An initial can be enlarged.

A formula introduces an entry.

Voynich paragraphs likewise have opening behaviour that differs from interior text.

The right question is not immediately:

This glyph means “begin”.

It is:

What kinds of forms gain probability when a new textual unit begins?

That question can be answered before meaning.

And any successful semantic model should eventually explain the answer.

Labels Are Not Tiny Paragraphs

The manuscript includes isolated word-like labels beside stars, diagram components and other illustrated features.

It is tempting to merge them into the same statistical population as running text.

That can be misleading.

A label may behave like a noun.

It may be an index code.

It may identify a component.

It may be a short instruction.

It may belong to a different register of the same writing system.

Whatever its meaning, its structural role differs from a token embedded inside a long paragraph.

Our retained work repeatedly became more reliable when structurally different textual environments were not casually pooled together.

The public lesson is simple:

same glyph system does not guarantee same textual job.

Currier A and Currier B Are Real Differences With Unknown Meaning

In the twentieth century, cryptanalyst Prescott Currier identified two broad varieties in Voynich text based on differences in character and word frequencies.

They became known as Currier A and Currier B.

This was an important advance because it showed that the manuscript is not statistically uniform.

But A and B are not translations.

They might eventually reflect:

different linguistic varieties;
different encoding regimes;
different subject matter;
different scribal habits;
different production periods;
different source exemplars;
or some combination.

They also interact with physical and visual distributions in complicated ways.

The careful sentence is therefore:

Voynich text contains large-scale statistical regimes conventionally called Currier A and B; their semantic cause remains unresolved.

That sentence keeps the discovery and the uncertainty together.

A/B Does Not Divide the Whole Manuscript Into Two Perfect Halves

Once a useful binary label exists, the mind wants to use it everywhere.

A or B.

One writer or the other.

One subject or the other.

One phase or the other.

The manuscript resists that simplification.

Some folios do not fit neatly.

Some textual properties vary continuously rather than categorically.

Some visual categories contain more than one textual profile.

Some physical gatherings cut across convenient descriptive boundaries.

So A/B is one useful projection of a higher-dimensional object.

It should not become the only coordinate system.

Local Vocabulary Is One of the Strongest Clues That Something Is Organised

If you pick up a chemistry textbook, certain words concentrate in certain chapters.

Atom.

Molecule.

Oxidation.

If you pick up a cooking book, another vocabulary concentrates.

Boil.

Slice.

Tablespoon.

Voynich word-like forms also show strong local and regional concentration.

This is one reason researchers have argued that the manuscript contains topical or semantic organisation.

That is plausible.

But local concentration can arise from more than topic.

a different source text;
a different cipher table;
a different writer;
a different formulaic template;
a different generation state;
a different page function.

So local vocabulary is evidence that the text has regional structure.

Calling the region “botany” is another inferential step.

Long-Range Structure Matters Because Randomness Has a Harder Time Faking Context

In 2013 Marcelo Montemurro and Damián Zanette analysed the distribution and co-occurrence of Voynich word-like forms. They reported long-range patterns compatible with the kind of topical structure seen in human language.

The finding matters because a text can have local regularity without having large-scale organisation.

You can generate a repetitive stream in which neighbouring strings look related.

It is harder to produce meaningful-looking regional concentration across a long document by naive random generation.

But again, the inference needs calibration.

Long-range statistical structure supports the proposition that the manuscript is organised.

It does not independently tell us what the organising variable is.

Topic is one candidate.

Production regime is another.

Cipher state is another.

The evidence narrows.

It does not name.

Network Statistics Also Say “Not a Simple Shuffle”

Also in 2013, Diego Amancio and colleagues examined Voynich text using complex-network and frequency measures.

Their results found that several statistical properties were compatible with natural-language texts and differed from shuffled versions of the manuscript.

This is often summarised online as:

Scientists proved Voynich is a language.

They did not.

The paper itself explicitly investigated statistical compatibility rather than decipherment.

The distinction matters because a test can reject one null model without identifying the true model.

If random shuffling fails, we have learned that order matters.

We have not yet learned whether that order comes from Italian, Latin, a cipher, a constructed notation or another constrained process.

“Language-Like” Is a Similarity Statement

This phrase deserves to be slowed down.

Suppose a feature of Voynich text falls inside the range observed for natural languages.

Then it is reasonable to say the feature is language-like.

But many systems can inherit or reproduce properties associated with language.

Ciphertext is generated from language.

Encoded notation may reflect linguistic structure.

Generated pseudo-text can be deliberately tuned to linguistic statistics.

A constrained copying process can preserve distributions.

So “language-like” should remain exactly what it says:

like language in this measured respect.

The next question is whether the same feature distinguishes natural language from the best alternative models.

The 2025 Naibbe Cipher Changes What a Statistic Can Rule Out

Michael Greshko’s 2025 Cryptologia study is important not because it claims to solve Voynich.

It does not.

Its importance is methodological.

Greshko constructed a historically plausible hand-operable verbose homophonic substitution cipher—the Naibbe cipher—and used it to encrypt Latin and Italian. The resulting ciphertexts reproduced several unusual statistical features associated with Voynich while remaining reversible to meaningful plaintext.

This is an existence proof.

It tells us that a class of hand-executable ciphers can produce more Voynich-like output than a simple substitution model would lead us to expect.

Therefore:

Voynich has property P.
Naive cipher X cannot produce P.
Therefore Voynich is not ciphertext.

is no longer a sufficient argument.

The cipher alternative has to be represented by stronger models.

A Cipher Can Preserve Structure Because It Starts With Structure

This idea is easier to understand with an ordinary example.

Suppose English uses “the” very frequently.

A simple substitution cipher might turn every “the” into the same three symbols.

The high frequency survives.

A more elaborate cipher may split “the” across several different encoded forms.

Some frequency signatures weaken.

Others survive at a higher structural level.

If the encoding expands common units, uses multiple substitutes or adds structural constraints, the ciphertext can look very different from ordinary plaintext while still inheriting non-random organisation.

This is why statistics must be compared against realistic alternatives.

A weak alternative makes the favourite theory look stronger than it is.

Natural Language Remains a Serious Hypothesis

None of this means the natural-language hypothesis has been defeated.

Far from it.

Claire Bowern and Luke Lindemann’s 2021 review in the Annual Review of Linguistics surveyed the manuscript through linguistic and computational evidence and argued that treating the text as language-like remains fruitful.

Natural-language models can account for:

frequency hierarchies;
recurrent word-like forms;
regional vocabulary;
co-occurrence patterns;
structural regularity;
and differences between textual varieties.

The question is not whether natural language can explain structure.

It can.

The question is whether the available evidence uniquely requires natural language in a recoverable form, and whether any proposed language can produce a reproducible reading.

That bridge has not yet been crossed.

Criticism Matters Because Good Reviews Need Adversaries

The Bowern and Lindemann review also attracted criticism from Torsten Timm and Andreas Schinner, who argued that some linguistic interpretations underweight alternative generation mechanisms and structural peculiarities.

This disagreement is useful for readers.

It reminds us that “the linguistic view” and “the generative view” are not merely two internet camps shouting at one another.

They make different claims about what the measured regularities require.

When serious researchers disagree, the student should ask:

Which observation do both sides accept?
Which inference differs?
Which alternative mechanism is being proposed?
What new measurement could distinguish the mechanisms?

That is a far more productive habit than asking which side sounds more confident.

“It Has Zipf’s Law” Is Not a Decipherment

Human language often displays a steep frequency relationship: a few words are extremely common, while many words are rare.

Voynich word-like forms also show non-uniform frequency behaviour.

This is sometimes connected to Zipf-like distributions.

Interesting?

Yes.

Unique to natural human language?

No.

Many generative and transformed systems can produce heavy-tailed frequency distributions.

The useful question is therefore not:

Does Voynich look Zipfian?

but:

Which combination of statistics is difficult for competing mechanisms to reproduce simultaneously?

A single famous statistic rarely carries an entire undeciphered writing system.

Entropy Is Useful Only If We Remember What Was Measured

Information theory gives researchers ways to quantify predictability.

If one character strongly predicts another, uncertainty decreases.

If many possibilities remain equally likely, uncertainty increases.

Voynich has unusual character-level predictability, and this has generated extensive debate.

But entropy depends on representation.

Change the glyph segmentation and the measured alphabet changes.

Change the transcription convention and some dependencies can move.

Merge compound forms and the sequence looks different.

So an entropy result should always travel with its unit definition.

Again:

measurement is not independent of representation.

Frequency Does Not Give You a Dictionary for Free

A common beginner move in cryptanalysis is to find the most frequent symbol and call it the most frequent letter of a guessed language.

For a simple monoalphabetic substitution cipher, this can be a useful first clue.

Voynich is not that obliging.

If the system uses homophones, verbose encoding, syllabic units, nulls, positional variants or multi-glyph mappings, raw frequency no longer maps cleanly onto plaintext letter frequency.

If the visible units are morphemes or syllables rather than letters, the comparison changes again.

If spaces do not equal word boundaries, word-frequency matching becomes even weaker.

Frequency is evidence.

A dictionary requires a model.

The model has to earn itself across large amounts of text.

A Proposed Translation Must Survive Unseen Text

This is the simplest strong standard for any claimed solution.

Do not judge the method only on the passage that inspired it.

Freeze the rules.

Then move elsewhere.

Can the same mapping read a page the solver did not use while inventing the key?

Can another researcher apply the rules without discretionary reinterpretation?

Does the output produce coherent grammar rather than isolated recognisable words?

Does it explain repeated forms consistently?

Does it handle Currier differences?

Does it account for labels and running text?

Does it explain positional effects?

Does the translation make falsifiable predictions about imagery?

If the rules must be reinvented on every page, the method is not yet a decipherment.

The Image Can Help a Translation—and Also Contaminate It

Suppose a page contains a plant.

A solver expects botanical words.

Then a token is decoded as “leaf”.

The image seems to confirm it.

But where did “leaf” come from?

If the image influenced the proposed key, and the proposed key is then validated by the same image, we have built a circle.

Images are valuable independent evidence only when the textual rule is sufficiently constrained before the image is used as confirmation.

A strong test would look more like this:

Derive the textual mapping from one dataset.
Freeze it.
Apply it to an unseen illustrated page.
Ask whether the resulting semantics predict something specific about the image.

Prediction is stronger than recognition after the fact.

Handwriting Variation and Textual Variation Need Separate Axes

Suppose two pages use different token frequencies.

Does that mean two people wrote them?

Not necessarily.

Suppose two pages look palaeographically different.

Does that mean the underlying language changed?

Not necessarily.

A writer can change register.

A writer can copy different source material.

Two writers can use the same textual system.

A cipher regime can change without a hand changing.

This is why the five-scribe debate matters beyond palaeography.

If we prematurely equate hand, language, topic and section, one uncertain classification propagates through the whole theory.

Keeping them separate feels less elegant.

It produces a stronger map.

The Text Is Structured Across More Than One Scale

One of the most durable outcomes of our retained work is that useful structure survives at several scales.

Glyph scale: some signs strongly prefer certain contexts.
Token scale: some word-like units recur and form families.
Line scale: beginnings and endings differ.
Paragraph scale: opening positions have distinctive behaviour.
Page scale: local vocabularies and layouts differ.
Regional scale: Currier and other distributions form large patterns.
Physical scale: some patterns interact with bifolia and gatherings.

The mistake would be to assume all seven scales encode the same underlying variable.

They might not.

A language has phonology, morphology, syntax, discourse and topic.

A cipher can add key state and operational conventions.

A manuscript adds physical construction.

Real information systems are layered.

Voynich appears to be no exception.

The Absence of Decipherment Does Not Mean the Statistics Failed

There is a common misunderstanding about research that does not reach the final answer.

If statistics have not decoded the manuscript, were they useless?

No.

They have changed the hypothesis space.

They have made simple randomness less plausible.

They have revealed local and long-range organisation.

They have shown that page position matters.

They have made uniform-text models less plausible.

They have created benchmarks that cipher and language hypotheses must meet.

A measurement does not fail because it does not produce a translation.

Sometimes its job is to tell us which translations are too easy.

The Hoax Hypothesis Also Has to Pay the Structural Bill

“It is meaningless” sounds like an escape from the decoding problem.

It is not.

A meaningless-text hypothesis has to explain why the text is structured as it is.

Why are some forms position-sensitive?

Why do regions differ?

Why do word-like families recur?

Why is there long-range organisation?

Why do labels and paragraph text interact with page architecture?

A hoax model can answer these questions if it includes a sufficiently rich generation procedure.

But once it includes that procedure, “meaningless” is no longer an explanation.

The procedure becomes the explanation.

And that procedure can be tested like any other model.

A Constructed Language Would Still Need Historical and Structural Evidence

Another possibility sometimes raised is that Voynichese is an invented language or notation.

This is logically possible.

But “constructed” does not solve anything by itself.

A constructed language still has:

an inventory;
combination rules;
semantic conventions;
a historical creator or community;
an intended use;
and usually some relationship to known linguistic or intellectual traditions.

The same burden remains.

Show the rules.

Apply them consistently.

Recover unseen material.

Explain why the physical and visual manuscript looks the way it does.

A hypothesis category is not a mechanism.

What Would Make a Textual Theory Strong?

Stable units. The method states how glyphs and compounds are segmented.
Stable rules. The mapping does not change opportunistically from page to page.
Coverage. It explains large amounts of text, not a few selected labels.
Prediction. It works on material withheld from the theory-building stage.
Grammar or mechanism. The outputs are generated by an intelligible system rather than free interpretation.
Position. It accounts for line and paragraph effects.
Regional variation. It explains Currier A/B and other non-uniformity.
Labels. It handles isolated labels and continuous text coherently.
Physical compatibility. It does not require impossible page order or chronology.
Visual prediction. Where images are used, semantics predict image properties rather than merely borrowing from them.
Historical plausibility. The language, cipher or notation can exist in the relevant historical world.
Replication. Independent researchers can apply the method and get comparable results.

A theory does not need to be perfect on day one.

It does need to become harder to change as evidence accumulates.

What We Can Say About the Writing With High Confidence

It uses a limited recurring inventory of sign forms under any reasonable transcription.
Those signs combine non-randomly.
Space-delimited forms recur with highly unequal frequencies.
Many forms belong to visibly related families.
Position in the line and paragraph affects distributions.
The text is not statistically uniform across the manuscript.
Currier A/B captures a major part of that variation.
Some word-like forms are strongly local.
Labels and running text create different contexts.
Long-range and co-occurrence structure is measurable.
Simple random shuffling is an inadequate model.
Simple substitution is not the only relevant cipher comparison.

Notice how much we have learned without assigning a single translation.

What We Still Cannot Say With High Confidence

Which language, if any, lies underneath.
What sound value a glyph has, if sound values are relevant.
Whether a glyph corresponds to a letter, syllable, morpheme, code group or something else.
Whether every visible space is a lexical boundary.
Whether frequent tokens are function words, code groups or generated forms.
Whether Currier A/B corresponds to language, topic, cipher regime, source, chronology or another variable.
Whether the text is plaintext, ciphertext, notation, constructed language, generated pseudo-text or a hybrid system.
Which proposed scribe count is ultimately correct.
What any complete sentence says.

This is the honest edge of the map.

Primary School: Find the Pattern Without Naming the Meaning

Give a young learner a made-up row of symbols:

△○○ △○□ △○○ ☆△○

Ask:

Which group repeats?
Which symbol is most common?
Which symbol appears only at the beginning?
Which groups look like relatives?

Then ask the important final question:

Do you now know what it means?

No.

But the child knows something about how it behaves.

That distinction is the beginning of serious pattern reasoning.

Lower Secondary: Turn Patterns Into Competing Models

For Secondary 1 and Secondary 2, take one positional pattern.

Suppose Symbol X appears mainly at line beginnings.

Ask students to generate at least four explanations:

it marks a new sentence;
it is decorative;
it records a cipher state;
the writer simply prefers that form at the margin.

Then ask what each explanation predicts elsewhere.

The exercise shifts the learner from:

I noticed a pattern

to:

I have several causal models that can now compete.

Upper Secondary: Learn the Null-Model Question

By Secondary 3 and Secondary 4, introduce a powerful question:

Compared with what?

If Voynich has repeated forms, compared with what?

If it has low entropy, compared with what?

If its network resembles language, compared with what?

A random shuffle?

A simple substitution cipher?

A sophisticated verbose cipher?

A generated pseudo-language?

The conclusion changes when the comparator changes.

This is one of the most important habits in statistics.

A pattern becomes evidence only relative to a model of what else could have produced it.

JC and Adult Readers: Ask for Identifiability

At a higher level, the central problem is identifiability.

If several different mechanisms can produce the same observed statistic, the statistic cannot identify the mechanism by itself.

Suppose natural language, verbose cipher and constrained generation all reproduce a token-frequency curve.

That curve is still informative.

It is not discriminating enough.

The next move is to search for another measurement on which the models diverge.

Line position.

Paragraph position.

Cross-page recurrence.

Label behaviour.

Physical gathering structure.

Image prediction.

The Voynich problem becomes much more sophisticated once the question changes from “does my model fit?” to “what observation would make my model uniquely preferable?”

A Parent and Teacher Guide

The manuscript is an unusually good way to teach students that pattern-recognition and explanation are different intellectual jobs.

Ask them to move through four sentences:

I observe… — describe the pattern without interpretation.
It might be caused by… — generate more than one model.
If that model is right, I should also see… — derive a prediction.
I would change my mind if… — state a falsifier.

Then insist on one more discipline:

Do not rename the pattern as the conclusion.

“Language-like” is not “language identified”.

“Frequent” is not “means the”.

“Currier B” is not “second author”.

“Label” is not “noun”.

That habit alone will improve how students read science, statistics and news.

Reader Checklist: Before You Believe a Voynich Translation

What transcription does the solver use?
How are ambiguous and compound glyphs handled?
Are visible spaces assumed to be words?
Were the decoding rules fixed before the showcased passage was chosen?
How much continuous text is successfully read?
Does the proposed language have coherent grammar?
Are common tokens translated consistently?
Does the method handle Currier A and B?
Does it explain line and paragraph positional effects?
Does it work on labels as well as running text?
Was image information used to invent the translation and then reused as validation?
Does the method succeed on unseen folios?
Can independent researchers reproduce it?
Are failures reported or quietly discarded?
Is the proposed historical cipher or language plausible for the manuscript’s material period?

A translation can be imaginative and still be worth reading.

It becomes a decipherment only when the rules outrun the imagination that created them.

Frequently Asked Questions

Does EVA translate Voynich?

No. EVA is a transcription convention that lets researchers represent visible glyph forms using convenient Roman characters. Those characters are labels, not decoded sound or meaning values.

Are Voynich spaces definitely word spaces?

No. Space-delimited units are useful for analysis, but their exact linguistic or encoding role is unresolved.

What are Currier A and B?

They are two broad statistical varieties of Voynich text identified through differences in character and word-like distributions. Their underlying cause remains unknown.

Do statistics prove Voynich is a natural language?

No. Several studies show strong language-compatible structure and reject naive random models, but alternative structured mechanisms can reproduce important features.

Does the Naibbe cipher solve the Voynich Manuscript?

No. It shows that a historically plausible class of hand-operable cipher can generate Voynich-like statistical properties from meaningful Latin or Italian plaintext, keeping complex-cipher hypotheses viable.

Why do line positions matter?

Because some glyphs and word-like forms occur with different probabilities at beginnings, endings and other structural positions. Any convincing generative or linguistic model should explain those differences.

Are the most common Voynich tokens function words?

Possibly, but not established. Frequency alone cannot determine whether a repeated form is a function word, morpheme, cipher group, structural marker or generated unit.

Could the text be meaningless?

It could in principle be generated without semantic plaintext, but any serious meaningless-text hypothesis must explain the manuscript’s multiscale regularities rather than merely call them gibberish.

What is the most important thing the writing tells us now?

That its producer followed constraints. The unresolved question is what kind of linguistic, cryptographic, notational or generative system created those constraints.

Research and Further Reading

The Final Idea

The writing in the Voynich Manuscript has spent a century inviting people to translate it too soon.

There is another way to listen.

Before the text tells us what it says, it tells us what it does.

It repeats.

It varies.

It respects position.

It forms local families.

It changes across regions.

It interacts with the page.

It is constrained.

That is not the answer.

It is the shape an answer will eventually have to fit.

Voynich | Everything eduKate Knows and Tested | What the Writing Does Before We Know What It Says