Voynich Drift Problem: Does Voynichese Change Gradually?

Voynich research loves boxes.

Currier A.

Currier B.

Herbal.

Zodiac.

Biological.

Pharmaceutical.

Stars.

Scribe 1.

Scribe 2.

These categories are useful because the manuscript really is heterogeneous.

But there is another possibility.

What if some of the boxes are snapshots cut from a moving system?

Imagine a colour gradient.

At the left, blue.

At the right, red.

Choose a threshold and we can classify every point as “blue” or “red”.

The categories are reproducible.

They may still be discretisations of a continuum.

This is the Voynich drift problem.

Currier’s A/B distinction remains one of the strongest discoveries in the manuscript’s textual history.

Yet newer work complicates the binary picture.

René Zandbergen’s 2024 extension introduces an RZ “C” language state for much astronomical and cosmological material, with statistics intermediate between traditional A and B.

Separately, Torsten Timm has argued that Voynich vocabulary and handwriting change continuously across manuscript order rather than breaking into sharply separated systems.

That second claim is controversial.

It is also scientifically useful because it asks a different question:

are Currier states different islands, or are they high-density regions along one changing landscape?

The answer matters for language, scribes, chronology, generation, page order and even how we validate a decipherment.

Quick Read

One-sentence answer: Currier A and B are robust statistical regimes, but current evidence does not require every Voynich page to belong to two perfectly discrete systems; newer RZ classification recognises an intermediate C-like state, while Timm’s 2026 continuous-evolution model argues that vocabulary and handwriting drift gradually—an important but contested hypothesis that must be tested against uncertain page order, bifolium production, document role and independent palaeography.

Currier A/B differences are real at character-group and whole-token levels.
Currier himself did not claim that A and B were proven distinct spoken languages.
Possible causes include language, dialect, topic, scribe, source, orthography or encoding state.
Currier did not classify every page.
Zandbergen’s newer RZ scheme introduces A, B and an intermediate C state.
RZ C covers much astronomical/cosmological and zodiac-related material previously left outside the binary classification.
In RZ’s current implementation, the frequency of EVA ed is one important discriminator, with C occupying intermediate values between traditional A and B.
RZ also proposes dialect-like subgroups within those larger classes.
The RZ scheme is explicitly still under development and should not be confused with decoded languages.
Timm’s 2026 model argues that vocabulary changes smoothly rather than at one abrupt A/B boundary.
He cites gradual folio-to-folio similarity decline and chains of intermediate word forms as support.
Timm connects this textual drift to a one-hand continuous-handwriting hypothesis.
That handwriting conclusion directly contests Lisa Fagin Davis’s influential multi-scribe model and is not consensus.
Current manuscript order may not be original order, which complicates any chronological “drift through the book” interpretation.
Bifolia are stronger production units than modern page adjacency, so drift should be measured across physical units as well as current folio sequence.
Document role can create apparent drift if the manuscript moves from herbal prose to diagrams to records.
A strong drift model must predict page properties beyond the ordering variable used to discover it.

The strongest safe conclusion is therefore neither “A and B are two languages” nor “A and B are an illusion”.

The manuscript contains reproducible textual states.

How sharply those states are separated remains an open modelling problem.

Currier A and B Are Too Strong to Explain Away

Any drift model begins with a fact it cannot simply dismiss.

Currier A and B differ.

Some character groups are common in one and rare in the other.

Some whole tokens change dramatically in frequency.

Herbal bifolia often remain internally coherent by regime.

The distinction has survived decades of independent observation.

Therefore continuous drift cannot mean:

there are no A/B differences.

It can mean something subtler:

A and B are useful clusters produced by a system whose underlying variables may change gradually.

Both statements can be true.

A temperature map can have cold and hot zones while temperature remains continuous.

The research question is whether Voynich has genuine discontinuities in its generating process or merely strong modes within a continuum.

Currier Never Proved Two Spoken Languages

The phrase “Currier languages” has created decades of semantic gravity.

Currier used A and B as labels for distinct statistical behaviour.

He did not decode them as French and Italian, or two unknown languages, or two dialects.

Possible explanations remain broad.

two languages;
one language in two dialects;
one language with different topics;
different scribes;
different exemplars;
different abbreviation habits;
different cipher or encoding settings.

This matters because a drift model can live inside several of those explanations.

A dialect can change gradually.

A scribe’s habits can change gradually.

An encoding process can change parameter frequencies gradually.

Topic mixture can change gradually.

“Continuous” describes the pattern of variation.

It does not identify the cause.

The RZ Classification Adds an Intermediate State

One of the most useful recent developments is Zandbergen’s extension of Currier classification.

The immediate problem was simple.

Currier left many astronomical, cosmological and related pages unclassified.

Those pages are not textually featureless.

When Zandbergen analysed them using quantitative character-group frequencies, much of the material occupied intermediate territory.

He therefore introduced RZ language C.

In the current scheme, one high-level discriminator is the frequency of EVA ed.

Traditional A pages have almost none.

B pages have much more.

Cosmological, astronomical, zodiac and some label groups fall between.

This is not proof of a smooth historical transition.

It is proof that a binary A/B map does not capture all textual variation equally well.

Intermediate Does Not Automatically Mean Transitional

This distinction is essential.

Suppose Group C has statistics halfway between A and B.

Three interpretations are possible.

Transition: the system moved historically from A through C toward B.
Mixture: C combines features from A and B because its document role requires both.
Third regime: C is a stable system with its own rules that happen to produce intermediate values on one statistic.

The word “intermediate” describes position in feature space.

It does not provide chronology.

To claim transition, we need ordered evidence.

Which state precedes which?

Does the same feature change monotonically across independently dated or ordered production units?

Without that, C expands classification resolution but does not prove drift.

Timm’s Continuous-Evolution Hypothesis Goes Further

Torsten Timm’s recent work proposes a stronger interpretation.

His model argues that Voynich vocabulary changes progressively through the manuscript.

He points to word networks in which many types connect through single-glyph edits.

He and Andreas Schinner have reported that similarity between folios can decline smoothly with distance rather than showing one clean step at the A/B boundary.

Timm also highlights chains of forms such as:

chol → cheol → cheo → chey → chedy

as examples of intermediates connecting familiar A-like and B-like families.

In his 2026 “One Hand, Five Labels” paper, he extends the same idea to handwriting and argues that both vocabulary and letterforms evolve continuously under one-scribe production.

This is a coherent model.

It is not settled consensus.

The One-Hand Conclusion Is a Separate Claim

Continuous vocabulary change does not logically require one scribe.

A workshop can share evolving conventions.

Several scribes can copy from source material that changes over time.

One master exemplar can be revised.

A changing cipher table can be used by several people.

Likewise one scribe can switch sharply between two established registers.

Therefore two propositions must remain separate:

Voynich textual features may vary continuously.
The manuscript may have been written by one continuously evolving hand.

Timm argues for both.

Lisa Fagin Davis’s 2020 palaeographic work argues for several distinct principal hands.

The Scribe Problem article owns that debate.

Here the important firewall is:

textual drift, handwriting drift and number of scribes are related variables, not interchangeable conclusions.

Current Page Order Is the Largest Chronology Trap

A drift hypothesis needs an axis.

Earlier → later.

But what is the Voynich chronological axis?

The present folio order is not guaranteed to be the production order.

Several bifolia appear to have been rearranged.

Missing leaves interrupt sequences.

Foldouts have complex histories.

The present binding may not be the first.

Therefore a smooth correlation with current folio distance can mean several things.

real production drift;
original-order remnants preserved imperfectly;
section/topic changes that happen to be ordered;
binding history creating pseudo-chronology.

This does not invalidate drift analysis.

It means chronology must be reconstructed rather than silently equated with page number.

The Bifolium Is the First Better Axis

The maker handled sheets.

Modern readers turn pages.

If drift is a production phenomenon, bifolia should be more coherent than arbitrary current neighbours.

This gives the hypothesis a strong test.

Measure the text state on all surfaces of one bifolium.

Then compare to adjacent bifolia.

Does the latent state change primarily between sheets?

Does it drift smoothly within a quire?

Does current adjacency matter less after physical sheet relationships are restored?

Currier’s original herbal observation—that bifolia tend to remain in one regime—is already evidence that production unit matters.

A continuous model must preserve that sheet-level coherence.

Document Role Can Mimic Drift

The manuscript is not one genre repeated for two hundred folios.

It moves among large plants, circular diagrams, human figures, pharmaceutical fragments and star-marked records.

Labels behave differently from prose.

Quire 20 behaves differently from herbal text.

If document roles are themselves ordered in the codex, vocabulary can change smoothly simply because the task changes smoothly.

For example, a transition from plant description toward astronomical annotation may gradually increase one character family and decrease another.

That is real distributional drift.

Its cause is genre, not time.

A strong chronological model must therefore compare like with like.

Herbal A against herbal B.

Running prose against running prose.

Labels against labels.

Otherwise content progression and temporal progression are confounded.

Topic Can Mimic Drift Too

Even within one document format, vocabulary changes with subject matter.

A herbal page about one plant family may use different words from another.

A medical recipe can differ by ingredient.

A star catalogue can move through coordinate systems.

Montemurro and Zanette’s long-range work suggests that token distributions are locally organised and correlate with broad manuscript regions.

That organisation is compatible with semantic content.

A drift model must therefore distinguish gradual change in topic from gradual change in writing mechanism.

One useful test is cross-topic recurrence.

Do the same function-like character shifts occur in several content regions?

If yes, mechanism drift strengthens.

If no, topical vocabulary may explain more.

A Scribe Can Drift Without Language Drifting

Handwriting changes over time.

Stroke length.

Slant.

Spacing.

Preferred allographs.

Ligature habits.

If one writer changes gradually, transcription can register that as textual drift even when the underlying language is constant.

A rare form becomes common.

One glyph variant replaces another.

A transcriber interprets the two as different characters.

Suddenly character frequencies drift.

This makes palaeographic normalisation crucial.

If drift disappears when plausible allographs are merged, some of it was scribal form rather than textual state.

If it survives, the case for deeper system change strengthens.

Several Scribes Can Create Apparent Discontinuities in a Continuous System

The opposite confounder matters too.

Suppose the underlying text convention changes gradually across production.

Scribe A handles one block.

Scribe B takes over later with different handwriting.

The palaeographic break may look sharper than the linguistic one.

Or several scribes may share two established orthographic conventions, producing strong A/B clusters even if production chronology is mixed.

This is why handwriting and text-state labels should be estimated separately and then compared.

If the same textual gradient crosses proposed hand boundaries, one-person explanations weaken.

If every major text transition aligns tightly with a hand transition, scribal convention becomes more plausible.

Cross-axis agreement is evidence.

A Cipher Can Drift

Cipher hypotheses are not limited to one frozen key.

A historical encoder can change tables.

Use different homophones.

Alter abbreviation conventions.

Adopt new code groups as work continues.

Therefore continuous surface change is compatible with cipher.

But a changing cipher creates a key-management problem.

How does the intended reader know which state applies?

Are states tied to page, section, scribe or physical packet?

A serious cipher-drift theory needs operational boundaries even if frequencies evolve gradually.

Otherwise the decoder cannot reproduce the intended plaintext.

A Natural Language Can Drift Too

Language is not stationary.

Dialect can shift.

Genre changes lexical preferences.

A scribe can alter spelling habits while copying the same language.

Technical vocabulary can accumulate across a compilation.

So continuous variation does not favour generation automatically.

The language model asks whether underlying grammatical classes remain stable while surface frequencies drift.

Function-like words may retain syntactic roles.

Morphological relationships may survive.

Pronunciation or spelling may vary systematically.

A decipherment should be able to map changing surface forms onto a stable enough source system for a reader to understand it.

A Generator Can Produce Drift Very Naturally

Copy-and-modify generation has a particularly natural drift mechanism.

Start with local tokens.

Copy them.

Modify one component.

Use recent forms as seeds for later forms.

The vocabulary performs a random walk through form space.

Near pages resemble one another.

Distant pages diverge.

Intermediate forms appear automatically.

This is one reason Timm interprets gradual vocabulary change within a generation framework.

The same model must still explain nonlocal constraints.

Currier regime coherence on bifolia;
label/prose register differences;
visual-section correlations;
long-range vocabulary organisation;
stable paragraph and line rules.

Drift is a strength of local generation.

It is not a complete explanation of the manuscript.

Asymmetry Between A-Like and B-Like Vocabulary Is Interesting

Timm’s work emphasises an asymmetry: forms typical of A can still appear in B, while strongly B-typical forms are less represented in A.

If robust, this resembles directional accumulation.

New families appear without older families vanishing completely.

That is compatible with evolution over time.

It is also compatible with nested document functions.

Suppose B adds grammatical or technical structures to an A-like base.

The subset relation can arise without chronology.

A directional vocabulary asymmetry is therefore valuable evidence, but its causal interpretation still needs external ordering.

Smooth Similarity Decline Needs a Null Model

Suppose folios close together in current order have similar vocabularies and similarity declines with distance.

That sounds like temporal drift.

But books are normally organised.

Nearby pages share topics.

Nearby pages share scribes.

Nearby pages share illustrations.

Nearby pages belong to the same quire.

Therefore a smooth distance-decay curve needs controls.

Compare with:

known multi-topic manuscripts;
known single-author texts;
texts arranged by subject;
randomised page orders preserving section blocks;
bifolium-aware permutations.

The question is not whether similarity decays.

It is whether it decays more smoothly than ordinary document organisation predicts.

The Original Order Problem Can Become a Test Rather Than a Limitation

If textual drift is real, it can help reconstruct page order.

Imagine two bifolia currently far apart but statistically adjacent along a learned drift trajectory.

A drift model predicts they were produced near one another.

Now inspect independent codicological evidence.

Paint transfer.

Quire structure.

Visual continuity.

Handwriting.

If those independent signals support the predicted adjacency, drift gains historical credibility.

If the model only reproduces current order, it may simply be learning the binder’s organisation.

This turns order uncertainty into a falsifiable opportunity.

Discrete and Continuous Models Need Not Be Enemies

Many real systems combine both.

A person’s handwriting drifts continuously.

Then they switch pens.

A dialect evolves gradually.

Then a different scribe takes over.

A cipher table accumulates modifications.

Then one new key is introduced abruptly.

Voynich may therefore contain:

continuous within-state drift;
discrete state changes;
document-role switches;
scribal switches;
binding reorderings.

Forcing the entire manuscript into “continuous” or “discrete” may be another false binary.

The better model can contain both if each is independently justified.

A Drift Model Should Predict Future-Looking Pages Without Using Their Position

The central validation idea is simple.

Learn a trajectory from some production units.

Remove page numbers and current positions from held-out bifolia.

Can the model place them correctly along the trajectory from text features alone?

Then compare that predicted order with independent codicology.

A real drift model should locate unseen sheets.

A descriptive smooth curve fitted after seeing every page order is much weaker.

Prediction converts continuity from visual impression into a historical hypothesis.

A Decipherment Must Explain the Drift

Suppose a future decipherment works.

It cannot ignore A/B/C-like variation.

If the source language is stable, surface differences need a mechanism.

dialect;
orthography;
abbreviation;
scribe;
cipher state;
genre.

If the source language itself evolves, the grammar and lexicon should change in historically plausible ways.

If the text is generated, the drift parameters should be explicit.

A solution that translates A one way and B another without explaining the transformation has split the problem rather than solved it.

The changing surface is part of the manuscript’s identity.

What Survives the Drift Work

Currier A and B are reproducible statistical regimes.
They are descriptive states, not proven separate spoken languages.
Many Voynich pages were not covered cleanly by the original binary classification.
RZ’s newer scheme identifies an intermediate C-like region, especially in astronomical/cosmological and zodiac material.
Intermediate feature values do not automatically prove temporal transition.
Timm’s recent work provides a coherent continuous-vocabulary-evolution hypothesis with reported gradual similarity changes and intermediate word forms.
Timm’s extension from textual drift to one-hand handwriting evolution is contested and not consensus.
Current page order is an imperfect proxy for production chronology.
Bifolium, document role, topic and proposed hand must be controlled.
Continuous and discrete processes can coexist.
A strong drift model should predict the placement or state of unseen production units.

What Does Not Survive as Established Knowledge

Currier A and B are proven two different languages.
Currier A and B are merely an illusion.
RZ C is proven a historical transition stage between A and B.
Intermediate statistics prove chronological order.
Current folio order equals production order.
Timm’s 2026 continuous-evolution model is settled fact.
Continuous vocabulary drift proves one scribe.
Continuous handwriting change disproves all multi-scribe interpretations.
A smooth similarity curve proves generated pseudo-text.
Every vocabulary difference reflects time rather than topic or genre.

The changing states survive.

The single causal story does not.

A Better Drift Analysis

Represent Currier A/B and newer intermediate classifications as measurements, not meanings.
Measure continuous features in addition to categorical labels.
Analyse at page, folio and bifolium levels.
Separate running prose, labels and other document roles.
Control proposed hand or repeat under several hand models.
Test sensitivity to allograph regularisation.
Do not assume current folio order is chronology.
Use codicology to constrain possible production orders.
Compare gradual and change-point models directly.
Penalise unnecessary state switches and unnecessary drift parameters alike.
Predict held-out bifolium states without their page numbers.
Require any causal explanation—language, scribe, cipher or generator—to reproduce the same trajectory.

What Would Count as a Real Drift Breakthrough?

Imagine a continuous text-state model is trained on a subset of bifolia without using current folio numbers.

It discovers a trajectory with A-like, intermediate and B-like regions.

Held-out sheets can be placed on that trajectory from text alone.

The predicted ordering differs from current binding in several places.

Independent codicology then supports those predicted rearrangements.

Palaeographic features change along the same trajectory but retain several discrete hand-specific offsets.

Document role explains some variance but not the underlying progression.

A compact mechanism—one evolving orthography, cipher state or production process—predicts the same pattern.

That would turn “the manuscript seems to drift” into a historically constrained production model.

A real drift breakthrough will not merely draw a smooth line through the current page order. It will use the line to predict where the physical manuscript itself has been rearranged.

Primary School: Categories Can Hide a Gradient

Make a row of colour cards from blue through purple to red.

Ask a child to divide them into “blue” and “red”.

The child can do it.

Now ask whether the purple cards suddenly jumped from blue to red.

No.

A useful category can sit on top of continuous change.

This is the basic difference between Currier classes and a drift model.

Lower Secondary: Change the Order

Create ten text samples that gradually change one letter frequency.

Put them in correct order.

The drift is obvious.

Shuffle the samples.

The same texts now look like several disconnected categories.

Students discover why unknown page order makes historical drift difficult to infer.

Upper Secondary: Topic or Time?

Prepare two sets of texts.

In Set A, vocabulary changes gradually because the author moves from plants to astronomy.

In Set B, the topic remains fixed but spelling habits gradually change.

Both show drift.

Ask which controls distinguish content drift from writing-system drift.

The exercise teaches why sequence alone cannot identify cause.

JC and Adult Readers: State-Space Versus Change Points

At a higher level, the problem can be represented with competing statistical models.

A discrete model assumes a small number of states and abrupt transitions.

A continuous model assigns each production unit a location in latent feature space and allows gradual movement.

A hybrid model allows both drift and occasional jumps.

The models should be compared on held-out prediction, not merely visual fit.

Crucially, production-order uncertainty should be integrated rather than ignored.

The best model is the one that explains observed variation with the least unsupported structure and makes correct predictions about independent manuscript evidence.

A Parent and Teacher Guide

Keep Currier A/B as strong descriptive evidence.
Do not turn “language A/B” into literal language identity.
Recognise newer intermediate classifications without making them chronology.
Separate continuous textual drift from one-scribe claims.
Control topic, document role and physical bifolium.
Never equate current page order automatically with production time.
Compare discrete, continuous and hybrid models.
Demand independent order predictions.

The general lesson is:

a clean category boundary can be scientifically useful without being a claim that reality itself changes discontinuously at that boundary.

Reader Checklist: Before You Claim Voynich Drift

Which feature is drifting?
Is the feature character-, word-, page- or hand-based?
What ordering variable defines “earlier” and “later”?
Is current folio order being assumed to be production order?
Are bifolium relationships preserved?
Are labels separated from prose?
Is topic controlled?
Is Currier/RZ state controlled?
Is proposed hand controlled?
Does allograph normalisation change the drift?
Does a change-point model fit equally well?
Does a hybrid drift-plus-jump model fit better?
Can the model place unseen bifolia without their page numbers?
Do independent codicological signals support the predicted trajectory?
Is a causal explanation being inferred beyond what continuity itself establishes?

Frequently Asked Questions

Are Currier A and B real?

Yes as statistical text regimes. Their cause and semantic status remain unresolved.

Are they proven two languages?

No. “Language A/B” is historical terminology for distinct textual statistics, not decoded linguistic identity.

What is RZ language C?

René Zandbergen’s recent extension classifies much astronomical, cosmological and zodiac-related material in an intermediate statistical state called C. It remains descriptive and under development.

Does C prove transition from A to B?

No. Intermediate feature values can reflect transition, mixture or a distinct third regime.

What does Timm argue?

His recent work argues that vocabulary similarity and handwriting forms evolve gradually through manuscript order, and he interprets this as consistent with one continuously evolving scribe and local copy-modify text generation.

Is that accepted?

No. It is a current contested hypothesis, especially where it challenges the influential multi-scribe palaeographic model.

Why does page order matter?

A temporal drift claim requires chronology, but current Voynich folio order may reflect rearrangement, missing leaves and rebinding. Physical production units provide safer constraints.

What is the strongest current conclusion?

Voynich textual variation is richer than a simple two-box model. Strong A/B states coexist with intermediate and possibly gradual variation, but the relative roles of chronology, document type, scribe and mechanism remain unresolved.

Research and Further Reading

The Final Idea

The Voynich Manuscript may not be a cabinet full of perfectly separate boxes.

It may be a landscape.

There are real mountains.