The Voynich Manuscript contains one of the strangest gifts an undeciphered text can give us.
It contains differences large enough to see before we know what they mean.
Some pages prefer one family of forms.
Other pages prefer another.
Certain endings become common in one group and nearly vanish in another.
Certain character combinations reverse their fortunes.
Whole word-like forms change frequency.
The differences are not confined to one isolated line.
They persist across pages.
In the 1970s, Prescott H. Currier gave these two broad statistical varieties names that have survived ever since:
Language A.
Language B.
The names were useful.
They were also dangerous.
Because once you call something a language, the mind starts behaving as though a language has already been found.
It has not.
Currier A and B are strong descriptions of textual behaviour. Their cause is still an open question.
That distinction is the entire article.
Quick Read
One-sentence answer: Currier A and Currier B are two major statistical regimes in Voynich writing, identified through systematic differences in glyph combinations, word-like forms and page distributions, but the distinction does not by itself prove two natural languages, two dialects, two scribes, two subjects or two chronological phases.
- Prescott Currier formalised the A/B distinction in work presented in the 1970s.
- His classification uses differences visible at several textual scales, including signs, sign groups and recurring word-like forms.
- Currier himself used the word “languages”, but the term is descriptive shorthand, not a decipherment.
- A and B are especially prominent in the plant-dominated pages.
- Currier observed that whole bifolia in the herbal material tended to belong to one textual variety rather than mixing A and B arbitrarily across the two halves of the same sheet.
- That physical correlation is important because it ties textual behaviour to manuscript production structure.
- A/B differences also appear outside the herbal pages, though not every page fits neatly into the original binary.
- Later cluster analyses confirm a strong A/B separation while also finding internal substructure, especially within B-like material.
- Recent independent work has proposed a third intermediate statistical regime for some astronomical and cosmological pages, illustrating that the original binary may be a useful projection rather than the final classification.
- Currier originally associated handwriting differences with textual differences, but later palaeographic work and current debate make it unsafe to equate A/B directly with two people.
- Topic may contribute to the distinction, but visual section and Currier regime do not align perfectly.
- Chronology may contribute, but A/B is not currently a proven before/after sequence.
- Cipher state, source exemplar, register, scribal habit, language variety and production workflow remain possible explanatory families.
- The best models must explain both the statistical separation and the places where the separation becomes fuzzy.
Currier’s discovery is powerful precisely because it does not need translation to be real.
Who Was Prescott Currier?
Prescott H. Currier was a cryptanalyst with professional experience in language and cryptology.
His Voynich work was presented in the 1970s and circulated through the research community surrounding the manuscript.
Currier did something deceptively simple.
He stopped treating the manuscript as though all its writing were statistically interchangeable.
He noticed that some pages behaved differently enough that the difference could be measured.
This is a major conceptual move.
If you average two different systems together, the average can hide both.
Imagine mixing Singapore English essays with chemistry laboratory reports and then calculating one vocabulary profile.
The combined statistics are real.
They may describe neither genre well.
Currier’s work warned Voynich research against exactly that kind of averaging.
Why the Word “Language” Needs Quotation Marks
Currier called the two varieties languages A and B.
That does not mean he had translated them into two known languages.
He was naming statistical populations.
Modern readers should therefore mentally translate:
Currier Language A
into:
the textual regime Currier labelled A.
And likewise for B.
Why be so careful?
Because the word “language” carries assumptions.
- speaker community;
- grammar;
- vocabulary;
- meaning;
- dialect relationships;
- historical development.
Currier A/B may eventually correspond to some of those things.
At present, the label tells us about distribution before it tells us about semantics.
A and B Differ at More Than One Scale
A robust classification becomes more interesting when it does not depend on one fragile feature.
Currier A/B is detectable through several kinds of difference.
Character-level differences
Some glyph forms and positional variants occur at different rates across the two regimes.
Bigram and sequence differences
Certain adjacent character combinations are much more characteristic of A or B.
Word-like form differences
Space-delimited strings that are common in A can become uncommon in B, and vice versa.
Page-level distributions
The features cohere strongly enough that pages can often be assigned to one regime based on their aggregate profile.
That multiscale agreement is why Currier’s distinction survived.
It is not one dramatic glyph.
It is a bundle of distributional behaviour.
One Famous Difference: Forms Ending in EVA -dy
In modern EVA transcription, word-like forms ending in -dy are strongly associated with Currier B compared with A.
This is the kind of feature that makes the distinction visible even to a non-specialist once it is pointed out.
But do not read -dy as a known linguistic suffix.
EVA is a transcription notation.
The ending might eventually correspond to morphology.
It might reflect an encoding rule.
It might mark a positional transformation.
It might be generated by a system whose units are not linguistic letters at all.
The important fact is distribution:
this family behaves very differently across A and B.
The semantic explanation comes later.
Other Character Groups Lean Toward A
Currier also identified groups conventionally rendered in modern transcription as forms like chol, chor and related sequences as much more characteristic of A than B.
Again, the exact transcription letters are merely labels.
What matters is that the same visible sign combinations shift strongly in frequency between page populations.
When several independent features point in the same direction, a cluster becomes hard to dismiss as noise.
This is why later computational clustering has repeatedly rediscovered a strong A/B distinction using different feature sets.
The names came from Currier.
The separation is not dependent on trusting Currier’s intuition.
The Herbal Pages Made the Difference Especially Visible
The plant-dominated pages were central to Currier’s observations.
This is useful because the visual job appears relatively stable at a broad level.
Plant-like image.
Text around or beside it.
Yet the textual statistics can differ sharply.
That immediately complicates one easy explanation:
The text differences are simply caused by entirely different visual subjects.
Within a broad visual family, A and B can both occur.
Topic may still vary at finer resolution.
Different plant classes could use different vocabulary.
But the simple equation:
visual section = textual language
is already too crude.
The Bifolium Result Is One of the Most Interesting Parts
Currier observed that in the herbal material, the A/B division respected bifolia unusually well.
Remember what a bifolium is.
One physical sheet is folded to produce two leaves and four page surfaces.
If both halves of a physical sheet tend to belong to the same textual regime, the pattern is not merely semantic.
It intersects production structure.
This creates several possible explanations.
- A single writer or team may have completed one bifolium under one set of habits.
- Source material may have been copied in physical batches.
- Different encoding regimes may have been assigned at sheet level.
- Different source exemplars may have been distributed by bifolium.
- Production chronology may have preserved regime consistency within physical units.
The observation does not select among them.
But it is extremely valuable because it ties abstract statistics back to the manufactured book.
Physical Correlation Does Not Automatically Mean Chronology
If A and B are segregated by physical production units, one tempting theory is:
A was written first.
B was written later.
Possible.
Not proven.
Two regimes can be contemporary.
A workshop can use two scribes at the same time.
One scribe can alternate between two source types.
A cipher can switch tables.
A compiler can copy different notebooks into different sheets.
A physical partition proves organisation at production scale.
Chronology requires an additional ordering signal.
Currier Also Talked About Hands
Currier believed handwriting differences correlated with A/B differences in parts of the manuscript.
This made a compelling package:
different handwriting + different statistics = different scribes using different languages.
That interpretation is elegant.
Later research makes the equivalence less secure.
Lisa Fagin Davis’s 2020 palaeographic work proposed five hands rather than Currier’s simpler hand structure.
A 2026 critique by Torsten Timm disputes whether five discrete hands are required at all.
The public-safe conclusion is therefore:
handwriting variation and A/B variation interact, but they should be kept on separate analytical axes until the relationship is established.
One uncertain map should not be used to label the other automatically.
Two Regimes Do Not Necessarily Mean Two People
One person can produce different statistical profiles.
Write a legal letter.
Then write a shopping list.
Your vocabulary changes.
Sentence length changes.
Abbreviations change.
Layout changes.
Same writer.
A medieval copyist could likewise produce different statistical regimes when:
- copying different sources;
- using different abbreviational conventions;
- switching registers;
- encoding with different procedures;
- working at different stages of a project.
Therefore A/B cannot be converted directly into a headcount.
It is evidence that production states differ.
The human explanation remains open.
Two Regimes Do Not Necessarily Mean Two Natural Languages
This is the most important semantic restraint.
Imagine an English text encrypted using two different code tables.
The plaintext language remains English.
The ciphertext distributions can differ.
Imagine one language written with two orthographic conventions.
Again, distributions differ.
Imagine two genres in the same language.
Again, vocabulary differs.
Imagine one generated system operating with two parameter states.
Again, output differs.
Therefore:
two statistical regimes ≠ two natural languages.
Two natural-language varieties remain one possible explanation.
They are not the only explanation.
Two Regimes Do Not Necessarily Mean Two Topics
Topic is attractive because topic changes vocabulary in ordinary language.
A medical chapter uses different words from an astronomy chapter.
So perhaps A and B simply encode different subjects.
The problem is that A/B cuts through broad visual categories.
Both regimes occur in plant-dominated material.
Some other regions are heavily skewed toward one regime.
Some pages are difficult to classify cleanly.
This leaves several possibilities:
- A/B reflects topic at a finer level than the visible section labels;
- A/B reflects something other than topic;
- topic contributes but is not the only variable.
The third possibility is often the most realistic in complex systems.
Several variables can shape one distribution.
The Biological Pages Make B Look Seductively Semantic
The human-figure and pool pages are strongly associated with B-like text.
This creates an appealing inference:
B is the language of bodies, bathing or medicine.
That is exactly the kind of inference we should preserve as a hypothesis rather than a fact.
Why?
Because B also occurs elsewhere.
And A/B may be influenced by production, author, source or encoding state in addition to subject.
A strong semantic interpretation would need to predict a specific textual feature from a specific visual or conceptual feature across pages.
Association is a clue.
Meaning requires discrimination.
B Is Not One Perfectly Uniform Thing
Later analyses have repeatedly found that B-like material contains internal structure.
This matters because a classification can be real and still hide subgroups.
“Mammal” is a real category.
It contains whales, bats and humans.
A/B can work the same way.
Strong top-level separation does not guarantee internal uniformity.
This becomes important when someone builds a theory from “B” as if every B page had identical behaviour.
Better questions are:
- Which B features are universal?
- Which belong to one subgroup?
- Which track visual section?
- Which track hand or layout?
- Which track physical gathering?
The more resolved our map becomes, the less useful a single binary may become for some questions.
Some Pages Sit Between the Old Labels
A classification system becomes most informative at its edges.
If every page were obviously A or B, the world would be simple.
Some astronomical and cosmological material has long been less comfortable inside Currier’s original binary.
Recent independent statistical work by René Zandbergen has proposed a third regime, labelled C, for some of this intermediate material.
This should not be mistaken for a new decoded “third language”.
It demonstrates something more methodologically interesting:
when a binary model leaves systematic residuals, the residuals may contain another dimension.
The original A/B distinction can remain valuable even if a later classification becomes richer.
This Is Why “A Versus B” Is Better Thought of as a Coordinate
Instead of imagining two sealed boxes, imagine a map.
One axis captures features strongly associated with A.
Another captures features strongly associated with B.
Pages occupy positions in that statistical space.
Many fall close to one pole.
Some are intermediate.
Some form subclusters.
This way of thinking avoids an unnecessary metaphysical question:
Is this page really A?
and replaces it with:
How A-like and B-like is this page under this feature set?
That is often a more faithful question for a continuous real-world system.
The Feature Set Changes the Map
Cluster by single characters and one geometry appears.
Cluster by bigrams and another may appear.
Cluster by whole tokens and another.
Add line position and the map changes.
Add paragraph opening behaviour and it changes again.
Add physical quire information and a new relationship becomes visible.
This does not mean the analysis is arbitrary.
It means different feature sets answer different questions.
A useful classification should be robust across reasonable representations or should clearly state which representation creates the result.
This is especially important in Voynich research because transcription itself already contains segmentation choices.
A/B Could Be Orthography
One explanatory family is orthographic variation.
The same underlying language can be written differently.
Think of British and American spelling.
Colour / color.
Centre / center.
Now imagine a much more radical manuscript convention involving abbreviation, glyph choice or syllabic representation.
Two surface systems could look statistically distinct while encoding closely related underlying material.
This hypothesis would predict systematic correspondences between A and B forms.
Finding those correspondences would be highly informative.
Not finding them would weaken a simple orthographic model.
A/B Could Be Cipher State
If Voynich writing encodes plaintext, A and B could reflect different encryption procedures.
Different key tables.
Different homophone assignments.
Different null conventions.
Different verbose encodings.
Different stages of an evolving cipher.
The 2025 Naibbe-cipher work matters here because it demonstrates that historically plausible hand-operable systems can produce surprisingly language-like and Voynich-like statistical output.
A cipher-state model would need to explain why regimes correlate with physical units and visual regions in the ways they do.
It would also need a historical mechanism for switching states.
Possible is not sufficient.
Mechanism matters.
A/B Could Be Source Exemplar
Imagine one compiler copying from two source traditions.
Source A uses one vocabulary or abbreviation system.
Source B uses another.
The compiler preserves enough of the source conventions that the final manuscript contains two statistical regimes.
This would be especially interesting if source material were allocated by bifolium or quire.
It could explain why physical units and textual profiles correlate without requiring two underlying natural languages.
But this model would need external or internal transmission evidence.
Different sources are not visible merely because two clusters exist.
The cluster creates the question.
It does not identify the source.
A/B Could Be Register or Genre
A single language can behave very differently by genre.
Prayer.
Recipe.
Glossary.
Commentary.
Catalogue.
Instruction.
Each produces different distributions.
If the Voynich sections perform different textual jobs, register could contribute to A/B.
The difficulty is again the herbal material, where broad page morphology is similar but A and B both occur.
A genre explanation may therefore need finer distinctions than “plant page”.
Perhaps one plant-page family is descriptive and another procedural.
Perhaps one records identification and another properties.
These are hypotheses.
The text is not yet readable enough to choose among them.
A/B Could Be a Moving Production System
Another possibility is that the system itself changed over time.
A writer learns.
A notation stabilises.
Abbreviations become habitual.
A cipher is streamlined.
Source material changes.
What looks like two regimes could be snapshots from different states of one evolving process.
This idea becomes especially relevant in the current debate over scribal hands.
If handwriting variation is more continuous than a five-person model suggests, then some textual variation might also be revisited as process evolution rather than categorical author difference.
Again, this is a research direction, not a conclusion.
The physical order would need to support the proposed evolution.
One Variable Is Probably Too Simple
The most seductive question is:
What does A/B mean?
Perhaps the better question is:
Which combination of hidden variables produces the A/B separation?
Real manuscript production can combine:
- scribe;
- source;
- topic;
- register;
- chronology;
- layout;
- cipher procedure;
- physical gathering;
- intended use.
A/B may be a visible projection of several of these at once.
This explains why one-variable theories often fit some pages brilliantly and then drift elsewhere.
The manuscript may not be inconsistent.
Our explanatory dimension may be too narrow.
A/B Is a Perfect Example of Why Clusters Are Not Meanings
Suppose a machine-learning algorithm divides a collection into two clusters.
The clustering is statistically strong.
What are the clusters?
Male and female writers?
Early and late texts?
Science and literature?
Formal and informal?
The algorithm does not know.
It knows that items separate under the chosen features.
Meaning comes from connecting the clusters to independent variables.
Currier A/B is exactly this kind of problem.
The separation is strong.
The label for the hidden cause remains open.
What a Good A/B Theory Must Explain
- Why characteristic glyph groups change frequency.
- Why whole word-like forms change frequency.
- Why the distinction is especially clear in some page families.
- Why both regimes appear in broadly herbal material.
- Why bifolium-level consistency appears in the herbal section.
- Why some non-herbal regions are heavily B-like.
- Why some astronomical/cosmological pages sit awkwardly between classic A and B.
- Why B-like material contains internal substructure.
- How handwriting variation relates—or does not relate—to the regimes.
- How the proposed mechanism fits the physical chronology.
- Whether the distinction predicts any visual, semantic or codicological property not used to create the classification.
A theory that explains only one bullet has explained a feature.
A theory that explains the whole list begins to explain the system.
What Would Count as a Breakthrough on A/B?
Imagine a decipherment in which A and B both resolve into the same underlying language under two consistent orthographic rules.
That would be powerful.
Imagine material evidence showing two production phases with A consistently earlier and B consistently later.
Also powerful.
Imagine a bilingual key revealing that A and B are two cipher tables over the same plaintext language.
Transformative.
Imagine independent palaeography securely assigning A and B to distinct writer communities while controlling for register and source.
Again, transformative.
Notice the common feature.
A breakthrough links the statistical regime to an independent historical or semantic variable.
It does not merely produce a prettier cluster plot.
Primary School: Two Boxes Without Naming Them
Give a child twenty made-up symbol strings.
Ten often contain ○□.
Ten often contain △☆.
Ask the child to sort them.
They can discover two groups without knowing what either group means.
Then ask:
What have you learned?
The correct answer is not:
Box 1 is animals and Box 2 is food.
Nothing established that.
The child learned:
there are two different pattern families.
That is Currier A/B in miniature.
Lower Secondary: Separate Cluster From Cause
For Secondary 1 and Secondary 2, give students two groups of English texts.
One group contains short messages.
The other contains formal essays.
Let them discover the statistical differences.
Then reveal that all texts were written by the same person.
The lesson becomes immediate.
A strong cluster can reflect genre rather than author.
Now reverse the exercise.
Give two authors the same genre.
Again, clusters can emerge.
The student learns that the same observed separation can have multiple hidden causes.
Upper Secondary: Build a Confounder Table
By Secondary 3 and Secondary 4, list the possible causes of A/B:
- scribe;
- topic;
- chronology;
- source;
- language;
- orthography;
- cipher state;
- genre.
Now make a second column:
What independent evidence would distinguish this cause?
Scribe → palaeography.
Chronology → physical sequencing or dated layers.
Topic → decoded semantics or predictive visual association.
Cipher state → recoverable transformation rules.
This is causal inference in a medieval manuscript.
JC and Adult Readers: Think of A/B as a Latent-Variable Problem
At a higher level, A/B is a classic latent-variable problem.
We observe outputs.
Glyph distributions.
Bigrams.
Tokens.
Page clusters.
We do not observe the hidden variable directly.
Was it author?
Topic?
Encoding state?
Several hidden variables?
The goal is not to choose the most narratively satisfying latent variable.
It is to find an observation that makes competing latent models produce different predictions.
That is the same logic used throughout modern science.
A Parent and Teacher Guide
Currier A/B is a wonderful way to teach students one difficult lesson:
classification is not explanation.
You can sort animals before understanding evolution.
You can cluster stars before understanding stellar physics.
You can identify two disease patterns before finding the pathogen.
You can discover A and B before deciphering Voynich.
Ask students:
- What feature created the groups?
- Would another feature create the same groups?
- What hidden cause might explain them?
- What other hidden cause could produce the same result?
- What independent evidence would distinguish the causes?
That sequence turns classification into reasoning.
Reader Checklist: Before You Use Currier A/B as an Explanation
- Am I using A/B as a statistical label or as a semantic claim?
- Which features define the classification?
- Is the page strongly A/B or intermediate?
- Does the feature set change the classification?
- Am I equating A/B with scribal hands without independent palaeography?
- Am I equating A/B with topic because of visual section?
- Am I assuming A is earlier and B later without chronological evidence?
- Does the theory explain the herbal bifolium pattern?
- Does it explain B substructure?
- Does it explain awkward astronomical/cosmological pages?
- Does it predict an independent variable not used in the clustering?
- Could the same separation arise from orthography, cipher state, source or register?
If a theory cannot answer those questions, A/B may be decorating the explanation rather than supporting it.
Frequently Asked Questions
What are Currier A and B?
They are two major statistical varieties of Voynich writing identified through systematic differences in sign combinations and word-like distributions.
Are they two decoded languages?
No. Currier’s word “language” is a historical label for the two textual regimes. Their semantic and linguistic cause remains unknown.
Did two scribes produce A and B?
Currier associated different hands with the regimes in parts of the manuscript, but later palaeographic models are more complicated and currently contested. A/B should not be converted directly into a two-scribe model.
Is A older than B?
Not established. Chronological evolution is one possible explanation, but the current evidence does not justify treating A→B as a proven production timeline.
Are all herbal pages A?
No. Broadly herbal pages contain both A and B, which is one reason simple topic-based explanations are insufficient.
Why do bifolia matter?
Because in Currier’s herbal observations, entire physical bifolia tended to share one regime. That connects textual statistics to how the manuscript was physically produced.
Is there a Currier C?
Currier’s original classification was A/B. Recent independent analysis has proposed an additional intermediate C-like regime for some pages. This is a modern statistical extension, not a decoded third language.
What is the strongest conclusion from A/B?
The manuscript’s text is not generated under one uniform statistical state. At least two major regimes exist, and any serious explanation of the writing must account for why.
Related eduKateSG Reading
- Voynich | Everything eduKate Knows and Tested | What the Writing Does Before We Know What It Says
- Voynich | Everything eduKate Knows and Tested | The Manuscript Before the Mystery
- Voynich | Everything eduKate Knows and Tested | What We Actually Know
Research and Further Reading
- Prescott Currier — Papers and Observations on the Voynich Manuscript
- René Zandbergen — The Currier Languages Revisited
- René Zandbergen — Voynich Manuscript Analysis Introduction and Updated Language Classification
- Bowern & Lindemann — The Linguistics of the Voynich Manuscript
- Lisa Fagin Davis — How Many Glyphs and How Many Scribes?
- Torsten Timm — One Hand, Five Labels: A Critical Examination of the Five-Scribe Hypothesis
- Michael Greshko — The Naibbe cipher
The Final Idea
Currier found two worlds in a book nobody could read.
That is extraordinary.
But the brilliance of the discovery is easy to ruin by naming the worlds too quickly.
Two languages?
Maybe.
Two scribes?
Maybe.
Two topics?
Maybe.
Two cipher states?
Maybe.
Several hidden variables projecting into one visible separation?
Also possible.
The strong thing we know is simpler:
the manuscript changes state.
The next great discovery will tell us what changed with it.