Voynich Entropy: Why Voynichese Is So Predictable

Imagine I show you the beginning of a word and ask you to guess the next letter.

In ordinary language, you often have several plausible choices.

Now imagine a writing system in which the first few symbols tell you surprisingly much about what can come next.

Certain symbols almost never appear in certain positions.

Some combinations are common.

Others are nearly forbidden.

The beginning of a line changes the possibilities.

The end of a line changes them again.

Currier A and Currier B prefer different families.

Word-like units cluster around a relatively narrow neighbourhood of permissible forms.

This is one of the deepest quantitative facts about Voynichese.

At the character level, it is unusually predictable.

Information theory gives us a way to describe that predictability.

Entropy.

The word sounds like disorder.

In this context, think of it more simply:

How uncertain are we about what comes next?

If many next symbols are equally plausible, uncertainty is high.

If only one or two symbols are likely, uncertainty is low.

Voynichese repeatedly gives the second experience.

That has led to some very strong claims.

It cannot be natural language.

It must be generated.

It must be ciphertext.

It must omit phonetic information.

Each is an explanation family.

None follows automatically from the measurement.

Low entropy tells us that the visible system is strongly constrained. It does not, by itself, tell us what created the constraints.

That distinction is the entire article.

Quick Read

One-sentence answer: Voynichese has unusually low conditional character entropy compared with a wide range of modern and historical natural-language texts, meaning its next glyph is often strongly constrained by preceding glyphs and structural position, but that predictability can potentially arise from unusual orthography, loss of phonemic distinctions, abbreviation, glyph composition, line-position conventions, encoding or constrained generation, so entropy is a powerful constraint on theories rather than a standalone decipherment.

Entropy measures uncertainty, not meaning.
Character entropy asks how unpredictable a symbol is when viewed by itself.
Conditional character entropy asks how much uncertainty remains after we know preceding context.
Voynichese is especially unusual in conditional character entropy: neighbouring characters restrict what can follow very strongly.
Lindemann and Bowern compared Voynich with large modern multilingual corpora and historical manuscript transcriptions and found its character-level predictability exceptional.
They also tested substantial changes in glyph composition and transcription representation; those manipulations did not make the Voynich profile simply converge on ordinary natural-language controls.
This does not prove Voynich has no underlying natural language. One interpretation is that its visible script may suppress distinctions that ordinary orthographies preserve.
Glyph segmentation remains important. Matlach, Janečková and Dostál argued in 2022 that some complex Voynich symbols may be compounds or ligatures, which would change alphabet size, word lengths and entropy calculations.
Even after representation questions are acknowledged, the broad fact remains: the visible writing is unusually constrained.
Strong token-internal slot restrictions reduce uncertainty.
Line and paragraph position reduce uncertainty further.
Word-family structure reduces the space of plausible forms.
Currier A/B changes which constraints dominate in different manuscript regions.
Natural language, encoded text and constrained generation can all produce predictability under different mechanisms.
Greshko’s 2025 Naibbe study demonstrates that a historically plausible hand-operable cipher can preserve meaningful Latin or Italian plaintext while reproducing many Voynich-like statistical properties, so “unusual statistics” cannot simply be equated with “not ciphertext”.
Low entropy does not prove hoax, gibberish, cipher or language.
The best theory must explain where the predictability comes from and reproduce it while also accounting for the rest of the manuscript.

The entropy result makes the mystery narrower.

It does not make the answer obvious.

Entropy Without the Mathematics

Suppose I put one of four cards face down.

Each is equally likely.

You are uncertain.

Now I tell you the card is red.

Your uncertainty shrinks.

Now I tell you it is a heart.

Your uncertainty shrinks again.

Information is what removes uncertainty.

Character entropy asks a comparable question about a text.

If I pick a random character, how difficult is it to guess?

Conditional entropy adds context.

If I tell you the previous one or two characters, how much uncertainty remains?

In a highly constrained system, context removes a lot of uncertainty.

That is the Voynich problem.

Unigram Entropy and Conditional Entropy Are Different Questions

A script can have many characters and still be predictable.

Imagine twenty symbols.

If all twenty can occur anywhere, the system has substantial local freedom.

Now imagine the same twenty symbols divided into strict slots.

five may begin a token;
six may occur in the middle;
four may occur near the end;
some pairs are forbidden;
one line position allows a special group.

The alphabet has not shrunk.

The next-character uncertainty has.

This is why conditional entropy is especially revealing for Voynich.

The unusual feature is not merely which symbols exist.

It is how strongly their neighbours and positions restrict them.

Voynichese Is Predictable Inside the Token

The word-family article showed that Voynich tokens occupy restricted structural neighbourhoods.

This is one direct source of low uncertainty.

Not every glyph can follow every other glyph.

Some components behave like openings.

Some behave like cores.

Some behave like endings.

Once you know which part of the token you are in, many future symbols become implausible.

This slot-like behaviour compresses uncertainty naturally.

The unresolved question is what the slots represent.

phonological positions;
morphemes;
abbreviations;
encoded groups;
production templates;
several layers at once.

Entropy tells us the slots are constraining.

It does not name the slots.

Bigram Restrictions Do Much of the Work

A bigram is simply a pair of adjacent symbols.

In Voynich, some pairs are common and many theoretically possible pairs are rare or absent.

That means seeing the first glyph provides strong information about the second.

If glyph X is almost always followed by one small set of glyphs, the next-character choice becomes easy to predict.

This is a major reason conditional entropy falls.

Again, the same bigram restriction can have several causes.

A language may forbid certain sound sequences.
An orthography may write particular letters only in compounds.
A ligature may have been split into component symbols by transcription.
An encoding algorithm may constrain legal groups.
A token generator may use predefined slots.

The pair statistics are powerful.

Their causal meaning remains open.

Glyph Composition Can Artificially Lower or Raise Entropy

Now return to EVA and segmentation.

Suppose a visible complex symbol is really one glyph.

A transcription splits it into A+B.

Because A is almost always followed by B, the bigram AB becomes extremely predictable.

Part of the low entropy was created by segmentation.

The reverse can also happen.

If two meaningful components are merged into one character, some dependencies disappear from the sequence.

This is why Lindemann and Bowern explicitly investigated glyph composition rather than treating one transcription alphabet as unquestionable reality.

Their result is important because substantial reasonable manipulations of glyph composition did not simply make the Voynich conditional-entropy profile ordinary.

Representation matters.

It does not make the whole phenomenon disappear.

The 2022 Ligature Argument Makes Representation Even More Important

Matlach, Janečková and Dostál examined unusual symbol behaviour and proposed that several graphically complex Voynich symbols may be compounds or ligatures built from simpler components.

If that interpretation is correct, the effective alphabet could be substantially smaller than an ordinary EVA inventory suggests.

That would change:

character frequencies;
bigram frequencies;
word lengths;
conditional entropy;
comparisons with natural languages and ciphers.

The paper goes farther and proposes an encoding interpretation.

That conclusion remains debated.

The reader-safe lesson is more durable:

entropy is measured over the units we decide to count, so the character model itself is part of the hypothesis.

A serious entropy result should therefore remain traceable to visible glyph decisions.

Yet Voynich Remains Unusually Constrained Across Reasonable Representations

This is the point at which two apparently opposing truths must be held together.

Truth one: entropy depends on transcription and glyph segmentation.

Truth two: the unusual predictability is not easily erased by reasonable changes in those choices.

Lindemann and Bowern built comparison corpora spanning hundreds of modern-language samples and transcribed historical texts and examined how script size, glyph composition, abbreviation and positional variants influence entropy.

The broad result was that Voynich remained unusually predictable relative to the comparison materials.

This makes low conditional entropy a real manuscript-level constraint rather than a fragile artefact of one EVA file.

What remains unresolved is the mechanism.

One Natural-Language Explanation: The Visible Script May Lose Distinctions

Lindemann and Bowern discuss an important possibility.

Perhaps the underlying language has more distinctions than the visible script represents.

Imagine English written without vowels.

Many different spoken forms collapse toward similar written structures.

Or imagine a notation that records only selected consonantal or syllabic features.

The visible alphabet becomes more constrained because information has been omitted or compressed.

This does not mean Voynich is an abjad or vowel-less script.

The example illustrates a mechanism:

low visible entropy can coexist with rich underlying language if the orthography deliberately fails to represent some distinctions.

This keeps natural language viable while giving it a real burden: identify the missing distinctions and show how the visible forms recover them.

Abbreviation Can Also Lower Visible Uncertainty

Medieval abbreviation systems compress frequently repeated material.

A common ending may be reduced to one sign.

Repeated formulae can acquire conventional shorthand.

A small visible inventory can therefore stand for a larger linguistic inventory.

This can increase predictability because the same shorthand structures recur.

But abbreviation only becomes explanatory when the expansions are constrained.

If one glyph may stand for any missing sequence needed to make a translation work, the theory gains too much freedom.

A real abbreviation model should recover repeated, historically plausible expansion rules.

Low entropy then becomes a consequence of a specific shorthand technology rather than a vague excuse.

A Cipher Can Also Produce Predictability

Simple substitution ciphers often preserve too much of the plaintext distribution to explain all Voynich properties comfortably.

That once encouraged strong arguments against ciphertext.

But “cipher” is a large mechanism class.

Verbose substitutions can expand one plaintext unit into several ciphertext symbols.

Homophones can distribute one plaintext value across multiple visible groups.

State-dependent rules can constrain legal sequences.

A historically plausible hand system can therefore generate highly structured ciphertext.

Greshko’s 2025 Naibbe study matters because it demonstrates this constructively rather than merely saying it is possible.

His verbose homophonic substitution system can be carried out using fifteenth-century-available materials and, when applied to Latin and Italian plaintext, reproduces many key Voynich-like statistics while remaining decryptable.

It does not reproduce every Voynich property.

It does not identify the actual Voynich mechanism.

Its epistemic effect is narrower and important:

ciphertext remains a live explanation family even when the output exhibits language-like or unusually constrained statistics.

A Generated System Can Be Extremely Predictable by Design

Generation models have an obvious advantage when explaining low entropy.

The generator itself can restrict the choices.

Choose one token template.

Allow only a few prefixes.

Allow a few cores.

Allow a few endings.

Modify neighbouring words by small edits.

Reset at line boundaries.

The resulting text will be highly predictable.

This makes entropy compatible with constrained generation.

But the generator inherits a larger burden.

Why does the text have long-range organisation?

Why do Currier regimes exist?

Why are labels coordinated with images?

What information—if any—is preserved through generation?

Low entropy is easy for a generator to produce.

The whole manuscript is harder.

Low Entropy Does Not Mean Meaningless Text

This misconception is worth removing completely.

A highly predictable signal can carry enormous meaning.

A DNA coding system is constrained.

A mathematical formula is constrained.

A legal form is constrained.

A musical notation system is constrained.

A language’s spelling rules constrain letters.

Predictability measures the size of the next-choice space under a representation.

It does not measure whether a human cared about the content.

Therefore:

low entropy ≠ meaningless.

Meaning must be established through other relationships.

Low Entropy Does Not Mean Hoax

A hoaxer can generate constrained text.

So can an honest scribe using a technical notation.

So can a cipher clerk.

So can a copyist using a highly abbreviated exemplar.

The entropy result has no access to intention.

It cannot distinguish deception from legitimate notation without other evidence.

This is why arguments that leap from “unusual text statistics” to “meaningless medieval hoax” are too large.

Intent is a historical claim.

Entropy is a distributional measurement.

Do not make one do the other’s job.

Low Entropy Does Not Mean Cipher

The same caution applies in the other direction.

A constrained visible sequence can be encoded.

It can also be direct notation.

Natural languages with restricted orthography can be predictable.

Abbreviation can compress variability.

Technical terminologies can repeat.

A cipher theory therefore needs a mechanism, not merely an entropy anomaly.

The mechanism should produce the observed entropy profile from plausible plaintext and also explain other features such as line effects, word families and Currier variation.

Naibbe raises cipher plausibility.

It does not replace the need for actual decipherment.

Low Entropy Does Not Mean “Not Language”

“Language” has several representational levels.

Spoken distinctions.

Phonemes.

Morphemes.

Orthographic signs.

Abbreviations.

Cipher groups.

Entropy measured at one level need not match entropy at another.

If the Voynich script encodes larger linguistic units than letters, comparing its glyph entropy directly with alphabetic letters can be misleading.

If it collapses phonemic contrasts, visible predictability rises even while the underlying language remains ordinary.

This is why the strongest linguistic interpretation of entropy is not:

Voynich cannot be language.

It is:

if Voynich encodes language directly, its orthographic representation is unusually constraining and needs explanation.

Language-Like Global Structure and Low Character Entropy Can Coexist

This apparent contradiction confuses many discussions.

Studies such as Montemurro and Zanette’s 2013 work found long-range word distribution compatible with meaningful topical organisation.

Amancio and colleagues found several network and statistical properties compatible with structured human text rather than shuffled controls.

Lindemann and Bowern, meanwhile, show that character placement is unusually predictable.

These results do not cancel one another.

A system can have:

strong long-range organisation;
local topical structure;
highly constrained character sequences.

A technical notation can do this.

An encoded meaningful text can do this.

A natural language under an unusual orthography can do this.

A sophisticated generated system might also do this.

The manuscript therefore resists single-statistic verdicts.

Line Architecture Reduces Entropy

The line article gives us another important source of predictability.

If the first word of a line favours one inventory, and line-final positions favour another, knowing line position narrows the next-symbol possibilities.

Paragraph openings narrow them differently.

So Voynich conditional structure exists at more than one scale:

inside the token;
inside the line;
at paragraph boundaries;
inside Currier regimes.

These interacting conditions can make the visible text far more predictable than a model that treats all characters as one stationary stream.

A future mechanism should therefore reproduce conditional entropy not only globally but by position.

Word Families Reduce Entropy Too

If tokens are drawn from dense families rather than assembled freely, the next character becomes easier to predict.

Suppose a core has only three legal endings.

Once the core is recognised, most of the alphabet is no longer plausible.

If one ending is strongly Currier-B-biased, knowing the textual regime narrows the choice further.

If a line-final position favours one of the three, uncertainty collapses again.

This illustrates a crucial point.

Voynich predictability may not come from one mysterious rule.

It may emerge from several interacting constraints.

The goal is to identify which constraints are fundamental and which are consequences of others.

Currier A and B Are Different Constraint States

Currier A and B are useful here because they show that the constraint system is not globally uniform.

Some character groups, endings and token families are common in one regime and rare in the other.

This means the probability of a next symbol depends partly on which regime the page belongs to.

A whole-manuscript entropy number averages across states.

That average can hide mechanism.

A serious model should ask:

Which constraints are shared by A and B?
Which change?
Can one transformation convert the constraint set of one regime into the other?
Does hand, source or page function explain the change better?

If the same underlying mechanism generates both regimes under two parameter states, that would be highly informative.

Scribal Hand Can Change Predictability Without Changing Language

If several scribes exist, personal orthographic habits can influence entropy.

One writer may use more variants.

Another may prefer a smaller inventory.

One may use terminal flourishes often.

Another may compress them.

These choices change character probabilities without changing underlying meaning.

That is why entropy comparisons by proposed hand are useful—but only if hand classification is independent enough from Currier regime and page type.

Otherwise we may think we are measuring scribal entropy when we are measuring topic or textual state.

Predictability Is Not the Same as Repetition

A text can repeat many words and still have relatively unconstrained character transitions.

Another can use many different words but build them from rigid slot templates.

Voynich does both interesting things:

it has repeated and near-repeated tokens;
its internal character sequencing is strongly constrained.

Do not collapse these into one observation.

Word-family density and character entropy are related but distinct.

A good mechanism should explain both and show whether one produces the other.

Predictability Is Not the Same as Compression Ratio Either

People sometimes use “compressible”, “repetitive” and “low entropy” as though they were interchangeable.

They are related but not identical.

A compression algorithm exploits many forms of regularity.

Character entropy measures uncertainty under a particular probability model and unit definition.

A manuscript can have local low entropy but substantial higher-level variety.

Or strong repeated phrases but a less constrained alphabet.

Precise terms matter because different mechanisms affect different measurements.

Voynich does not become better understood when every kind of structure is renamed “low entropy”.

One Number Is Not Enough

A global entropy value is useful as a summary.

It can also hide the reasons for the summary.

Better questions include:

entropy by Currier regime;
entropy by proposed hand;
entropy at line beginnings versus middles;
entropy inside labels versus running text;
entropy inside different token positions;
entropy under alternative glyph segmentations;
entropy before and after proposed ligature decomposition.

Each conditional view can identify which variable removes uncertainty.

The question becomes causal:

what information lets us predict the next symbol?

That is much more informative than merely announcing that the manuscript is predictable.

Controls Determine What “Unusual” Means

Voynich can look astonishingly strange if compared only with modern English prose.

That is not a sufficient control.

A fair comparison should consider:

historical spelling variation;
manuscript abbreviation;
different script inventories;
languages with different morphology;
genre;
text length;
transcription conventions.

This is one strength of the Lindemann–Bowern work: it did not rely only on present-day English, but used modern multilingual and historical corpora while explicitly examining representational factors.

No control set is perfect.

But better controls make “unusual” a more meaningful word.

A Strong Theory Should Generate the Entropy, Not Merely Fit It

Suppose a model says Voynich is an abbreviated Romance language.

Good.

Take plausible Romance plaintext.

Apply the proposed abbreviation rules.

Does the output produce Voynich-like conditional entropy?

Suppose another model says the text is encoded.

Apply the encoding to plausible plaintext.

Does the output reproduce the entropy, word families and line effects together?

Suppose a generation model is proposed.

Generate a large corpus under fixed rules.

Does it reproduce not merely low entropy but also Currier-like variation and image-correlated structure?

This is the difference between:

my theory is compatible with low entropy

and:

my mechanism actually produces the observed constraint profile.

The Best Mechanism Should Explain Several Statistics at Once

A model that reproduces one entropy number can still be wrong.

The manuscript has a constellation of properties.

low conditional character entropy;
word-family density;
strong positional glyph restrictions;
line and paragraph effects;
Currier A/B variation;
local vocabulary;
long-range organisation;
integration with visual and document structure.

A mechanism that explains several independently is more compelling than one that is tuned to one statistic.

This is why Naibbe is methodologically interesting: it aims to reproduce many properties simultaneously rather than one selected metric.

It also has limitations.

Those limitations matter because a model should be judged on the whole profile, not only its successes.

Entropy Can Help Find the Right Unit of Writing

One of the most interesting uses of entropy is not deciding “language or cipher”.

It is testing what the units might be.

If splitting a complex glyph into components makes the system more regular in a principled way, perhaps the components are real.

If merging two consistently linked components removes an artificial bigram dependency, perhaps the compound is one grapheme.

If one segmentation aligns the corpus with historically plausible orthographic behaviour and predicts unseen forms, it gains support.

Entropy can therefore participate in representation discovery.

It should not dictate representation alone because many segmentations can optimise one statistic.

Palaeography must remain connected to the numbers.

Entropy Can Help Test Proposed Decipherments

Suppose somebody proposes a plaintext and encoding scheme.

The method produces convincing English translations.

Now reconstruct the visible text from the proposed plaintext under the claimed mechanism.

Does it reproduce the observed predictability?

If not, something is missing.

The proposed semantics may fit selected examples while the generative mechanism fails the corpus.

This is why entropy is a useful validation layer.

It is difficult for a translation to negotiate with a manuscript-wide statistic after the fact.

A real solution should make the low entropy less mysterious rather than asking us to ignore it.

What Survives the Entropy Work

Voynichese is highly predictable at the character level.
Conditional character entropy is unusually low relative to broad modern and historical comparison corpora.
Strong token-internal positional restrictions contribute substantially to the effect.
Line and paragraph positions add another layer of constraint.
Word-family structure narrows the space of legal next symbols.
Currier A/B shows that constraint states vary across the manuscript.
Transcription and glyph composition affect entropy and must remain explicit.
Reasonable representation changes do not trivially erase the broad unusual predictability.
Natural-language explanations remain possible if the visible orthography compresses or omits distinctions.
Abbreviation remains a plausible contributor under constrained historical rules.
Ciphertext remains viable, especially after constructive models such as Naibbe demonstrate that historically plausible encodings can reproduce many unusual properties.
Constrained generation remains viable at the surface level.
Entropy alone does not determine semantic status, intention or provenance.

The entropy result is one of the strongest negative constraints in Voynich research.

Whatever the manuscript is, it is not written with the local freedom of an ordinary alphabetic plaintext sampled naively from many familiar languages.

Something constrains it more strongly.

What Does Not Survive as Established Knowledge

A proof that Voynich has no underlying natural language.
A proof that the text is ciphertext.
A proof that the text is mechanically generated.
A proof that generated text is meaningless.
A proof that the manuscript is a hoax.
A proof that EVA identifies the correct character inventory.
A proof that proposed ligatures in every case are true compounds.
A proof that one entropy value identifies the writing system.
A proof that Currier A/B are two entropy-defined languages.
A proof that low entropy alone can reveal purpose or provenance.

The measurement survives.

The grand semantic verdict does not.

A Better Entropy Analysis

Name the transcription and alphabet.
State how compounds and ligatures are treated.
Separate unigram from conditional entropy.
Condition on token position.
Condition on line and paragraph position.
Separate Currier regimes.
Where possible, separate proposed hands and text roles.
Use historically and linguistically reasonable control corpora.
Test sensitivity to alternative reasonable representations.
Require explanatory models to generate the measured profile rather than merely describe it.

The question is not simply “how low is the entropy?”

It is:

which layers of structure remove the uncertainty?

What Would Count as a Real Entropy Breakthrough?

Imagine a new orthographic model proposes that several visible glyphs are compounds and that the system omits one class of phonemic distinction.

The model is derived from palaeographic evidence.

When applied consistently, it transforms Voynich character predictability into the range expected for a specific historical language family.

The same representation predicts morphological families and deciphers unseen text.

That would be a major breakthrough.

Or imagine a historically plausible encoding system, fixed in advance, produces Voynich-level entropy, line effects, word-family structure and Currier-like state changes from readable plaintext and then independently deciphers substantial manuscript passages.

That too would be transformative.

Or imagine a generation model reproduces the full statistical profile but repeatedly fails all tests of semantic cross-reference, while a language model succeeds.

The entropy data would then have helped distinguish mechanism rather than merely label strangeness.

The breakthrough comes when predictability stops being an anomaly and becomes an expected consequence of one independently supported system.

Primary School: Guess the Next Symbol

Create two simple games.

Game A:

A can be followed by B, C, D, E or F.

Game B:

A can only be followed by B.

Ask which game is easier to predict.

Game B.

That is the intuitive idea behind lower conditional entropy.

Then ask:

Does an easier-to-predict game have less meaning?

Not necessarily.

The child has learned the key distinction between predictability and semantics.

Lower Secondary: Same Alphabet, Different Rules

Give students the same six symbols for two invented writing systems.

System 1 allows any symbol after any other.

System 2 allows only twelve specific pairs.

Both have the same alphabet size.

System 2 is easier to predict because its transitions are constrained.

This teaches why alphabet size and conditional entropy are different.

It also shows why a Voynich alphabet debate cannot by itself settle the predictability problem.

Upper Secondary: Change the Segmentation

Give students a repeated compound written as XY.

In Dataset A, X and Y are separate characters.

In Dataset B, XY is one compound symbol.

Ask what happens to the predictability of Y after X.

In Dataset A, it is extremely predictable.

In Dataset B, the pair no longer exists as a transition.

The underlying ink is identical.

The representation changed.

This is a direct classroom model of the Voynich glyph-composition problem.

JC and Adult Readers: Think in Conditional Information

At a higher level, the useful object is not a single entropy value but a hierarchy of conditional uncertainty.

How uncertain is the next character given the previous character?

Given token position?

Given line position?

Given Currier regime?

Given proposed scribal hand?

Every conditioning variable that sharply reduces uncertainty points toward a structural rule worth explaining.

The mature question becomes causal:

which hidden production or linguistic variable makes this information predictive?

That is where information theory becomes manuscript history.

A Parent and Teacher Guide

Explain entropy as uncertainty. Avoid treating it as a synonym for disorder.
Separate alphabet size from transition freedom.
Ask which context makes the next symbol easier to guess.
Remember representation. One glyph versus two changes the statistic.
Do not confuse predictability with meaninglessness.
Compare mechanisms. Language, abbreviation, encoding and generation can all constrain output.
Require mechanisms to reproduce several observations together.
Use entropy as a validation layer, not a one-number verdict.

The transferable lesson is larger than Voynich:

a measurement can be highly informative without uniquely identifying its cause.

Reader Checklist: Before You Use Entropy to Declare What Voynich Is

Which transcription and alphabet were used?
How are ligatures and compounds treated?
Are you discussing unigram entropy or conditional entropy?
What historical and modern control texts were used?
Does the result survive reasonable segmentation changes?
How much predictability comes from token position?
How much comes from line or paragraph position?
Does Currier regime change the result?
Could scribal conventions explain part of it?
Could information be omitted by the visible orthography?
Could historically plausible abbreviation produce the constraint?
Could an encoded system reproduce the same profile?
Could constrained generation reproduce it?
Does the proposed mechanism also explain word families and long-range organisation?
What evidence would distinguish the competing mechanisms?

Frequently Asked Questions

What is entropy in Voynich research?

It is an information-theoretic measure of uncertainty in the symbol sequence. Conditional character entropy measures how much uncertainty remains about the next character after some preceding context is known.

Why is Voynich entropy unusual?

Its character placement is highly constrained, so preceding symbols and structural position make the next symbol unusually predictable compared with broad modern and historical language controls.

Does low entropy prove the manuscript is meaningless?

No. Predictability measures structural uncertainty, not semantic content or human intention.

Does it prove a cipher?

No. Ciphertext is one viable explanation family, but unusual orthography, abbreviation and constrained generation can also lower visible entropy. A cipher mechanism must be demonstrated.

Does low entropy rule out natural language?

No. It makes ordinary direct alphabetic plaintext a harder fit. A natural-language model may remain viable if the visible system omits distinctions, uses heavy abbreviation or encodes larger linguistic units.

Why do ligatures matter?

If a transcription splits one compound glyph into two characters, it creates an extremely predictable character pair. Merging or decomposing glyphs therefore changes entropy calculations.

What did Lindemann and Bowern find?

Their comparative work found Voynichese unusually predictable at the character level relative to extensive modern and historical corpora, and substantial reasonable changes in glyph composition were insufficient to make its conditional entropy simply ordinary.

What did the 2022 PLOS study add?

It explored unusual symbol-repetition behaviour and proposed that several complex characters may be ligatures or compounds, illustrating how character inventory and segmentation can alter quantitative conclusions. Its stronger encoding interpretation remains one debated hypothesis.

What does the Naibbe cipher contribute?

It demonstrates that a historically plausible hand-operable verbose homophonic substitution system can encrypt meaningful Latin or Italian while reproducing many key Voynich-like statistical properties. It keeps the ciphertext hypothesis viable without identifying the actual Voynich system.

What is the strongest conclusion?

The visible Voynich writing system is unusually constrained. Any serious explanation must identify a mechanism that naturally produces that predictability while also accounting for the manuscript’s word families, positional effects, text regimes, images and physical structure.

Research and Further Reading

The Final Idea

The Voynich Manuscript is not difficult because its writing is chaotic.

In one important sense, it is difficult because its writing is too orderly.

The next character has fewer choices than we expect.

Word shapes live in narrow families.

Line boundaries change the system.

Currier regimes change it again.

That orderliness is a real clue.

But it is not a label.

Language can be constrained.

Abbreviation can be constrained.

Ciphertext can be constrained.

Generated notation can be constrained.

The winning explanation will not be the one that points at low entropy and announces what Voynich must be.

It will be the one that can say:

This is the mechanism. These are the constraints it creates. This is why the next symbol becomes predictable. And these are the new pages it lets us predict before we read them.

At that point, entropy will have stopped being a mystery statistic.

It will have become a consequence of understanding.

Continue Through the Voynich Research Map

This article is one specialist node in eduKateSG’s larger Voynich research library. Return to the canonical master to see the physical, visual, textual and historical evidence in one continuous argument.

Structure is evidence; structure is not translation. The master keeps resemblance, statistical structure, historical possibility and decipherment claims at separate evidentiary levels.

Voynich | Everything eduKate Knows and Tested | Why Voynichese Is So Predictable: Entropy, Constraints and What They Do Not Prove