A text can have a direction before we know what it says.
Most Voynich prose is visibly written from left to right. The pen moves across the line that way. Paragraphs begin on the left. Spaces divide the stream. For years, that physical direction has been one of the manuscript’s least controversial features.
Then statistics asked a different question.
Not:
Which way did the scribe move?
But:
Which direction makes the visible sequence more predictable?
In 2026, Christophe Parisel reported a striking result in a preprint analysis of the Voynich text. At the character level inside token-like units, the sequence appeared more predictable when modelled in the reverse direction. Yet at the boundary between tokens, the dependency pointed the other way.
Inside the token: one directional preference.
Across the boundary: another.
That is the directional dissociation problem.
The most interesting possibility is not that Voynich “reads backwards”. It is that different structural layers may obey different directional rules.
Quick Read
- The ordinary visual writing direction of Voynich prose is left-to-right.
- A 2026 preprint by Christophe Parisel reports a different statistical phenomenon: token-internal character sequences are more predictable under right-to-left modelling, while cross-token boundary dependencies favour left-to-right structure.
- The paper calls this a directional dissociation between two structural layers.
- The same pattern was not observed in the four comparison languages used in the study: English, French, Hebrew and Arabic.
- The result is statistical, not semantic. It does not establish that the manuscript should be read right-to-left.
- It does not prove a cipher. The authors describe the pattern as compatible with cipher-like constraints and difficult for the tested simple generators to reproduce.
- The tested generators included a parametric slot model and a Cardan-grille implementation inspired by Rugg’s hypothesis; neither matched the full joint signature across the tested parameter space.
- The result depends on transcription, tokenisation and modelling assumptions, so replication under alternative representations matters.
- The most useful interpretation is architectural: the rules that form a token may differ from the rules that connect one token to the next.
- A successful Voynich mechanism should explain both layers simultaneously rather than fitting one and ignoring the other.
Physical Direction and Statistical Direction Are Not the Same Thing
This distinction is the first firewall.
If I write the English word because from left to right, the physical direction of writing is obvious.
But a statistical model can ask whether the letters are easier to predict from left-to-right or from right-to-left.
Those are different questions.
A suffix-rich language can make endings highly informative. A code can place strong constraints near one edge. A transformation can make the final part of a visible token determine earlier-looking structure more strongly than the reverse.
None of that requires the historical reader to reverse the text.
The existing Which Way Does Voynich Read? article owns the physical and geometric direction question.
This article owns something narrower:
directionality inside the statistical architecture of the visible text.
What Does “More Predictable Right-to-Left” Mean?
Imagine a sequence of characters inside thousands of Voynich tokens.
Train one model to predict the next character from the characters already seen when moving left-to-right.
Train a comparable model in the opposite direction.
If the reverse-direction model needs fewer bits on average to predict the sequence, then the visible internal structure is more constrained in that direction.
That does not mean the token “means more” backwards.
It means the statistical dependencies inside the token are asymmetrical.
This is plausible in many systems.
- A suffix can constrain what usually appears before it.
- A templated code can fill final slots under stricter rules than initial slots.
- A verbose cipher can build visible groups from one edge under a transformation procedure.
- An abbreviation system can attach endings according to preceding structural classes.
Directionality is therefore a property to explain, not a language label.
Then the Boundary Does Something Different
If Voynich were governed by one simple directional process, we might expect the same preferred direction to dominate everywhere.
The 2026 result instead reports a split.
Within tokens, one asymmetry appears.
Across the end of one token and beginning of the next, another dependency appears in the normal left-to-right sequence.
This matters because a token boundary becomes more than whitespace.
The boundary may participate in a higher-level rule.
The machine that builds a token may not be the same machine that chooses what can follow it.
Why That Is Hard for One-Layer Explanations
Suppose a model says every Voynich token is generated independently from a fixed slot template.
That model can create beautiful internal regularity.
But if tokens are chosen independently, it has no natural reason to produce a strong directional signature across token boundaries.
Now suppose a model says every next token is created only by copying and modifying the previous token.
That can create cross-token dependence.
But it must also reproduce the specific internal directional asymmetry rather than merely create local resemblance.
This is what makes a joint criterion stronger than one successful statistic.
A model does not win because it reproduces entropy.
Or Zipf.
Or word length.
Or boundary dependence.
It must reproduce the combination.
What Parisel Tested
The preprint compared Voynich with several natural-language controls and evaluated structured generators against a four-signature joint criterion.
The comparison languages included English, French, Hebrew and Arabic.
The tested generative families included:
- a parametric slot-based generator;
- a Cardan-grille generator representing a Rugg-style pseudo-text mechanism.
Across the tested parameter spaces, neither reproduced all four target signatures simultaneously.
That is meaningful.
It is also bounded.
The result does not eliminate all possible generators.
It eliminates or weakens the tested versions under the tested criteria.
This is exactly how a good null-model result should be stated.
Why “Cipher-Like” Is Not “Cipher Proven”
A cipher can naturally create layered directional constraints.
Plaintext may be read left-to-right.
A substitution table may generate several visible symbols per plaintext unit.
Those visible symbols can obey internal formatting rules.
The next plaintext unit can then influence the next visible group.
Two directions of statistical dependency can therefore emerge from one layered encoding system.
But other systems can also create layers.
- abbreviation;
- templated notation;
- morphology plus scribal formatting;
- multi-stage generation;
- hybrid cipher-and-abbreviation schemes.
The phrase “cipher-like” should therefore be understood as a structural comparison, not a verdict.
The canonical mechanism comparison remains Cipher, Plaintext or Generated System?.
Could Natural Language Produce This?
Possibly through representation effects.
An underlying natural language can be transformed before it becomes visible.
- Vowels can be suppressed.
- Common sequences can be abbreviated.
- One plaintext unit can expand into several visible glyphs.
- Spaces can be moved or graded.
- Affixes can become more regular than ordinary spelling.
So comparison with ordinary orthographic English or Latin is not a direct comparison with every possible encoded natural-language system.
The question becomes:
What transformation would produce this directional split while remaining historically and mechanically plausible?
Transcription Can Manufacture Direction If We Let It
This is the most important technical caution.
Voynich is analysed through transliteration.
If one compound glyph is represented as two EVA characters, a strong internal bigram is created automatically.
If a bench is split differently, the directionality changes.
If uncertain spaces are accepted as full word boundaries, cross-boundary statistics change.
If rare ligatures are normalised inconsistently, one edge can become artificially constrained.
Therefore the strongest replication should survive:
- more than one transcription;
- compound-glyph fusion;
- uncertain-space sensitivity analysis;
- Currier A/B separation;
- quire-level resampling;
- scribe/hand controls where feasible.
The EVA and Segmentation Problem is therefore not a side issue. It is part of the result’s foundation.
The Boundary May Be a Different Kind of Unit
Recent 2026 work by Liudmila Rozanova and Alexander Temerev pushes the boundary issue even further in a separate preprint.
Their analysis argues that ordinary assumptions—glyph equals letter, token equals word, blank equals word space—do not survive their matched-control tests cleanly.
They report unusually strong coupling between token-edge glyphs even though exact whole-token identity predicts little of the next token.
They also report that uncertain spaces behave more like internal junctures than confident spaces.
That does not settle the matter.
It does make one thing increasingly difficult to defend:
the idea that all of Voynich’s structure lives at one clean linguistic level.
A Layered Mechanism Has to Pay Rent at Every Layer
Suppose a model proposes three levels.
- Level 1 generates internal token structure.
- Level 2 chooses token-edge forms.
- Level 3 chooses document-level sequence.
That sounds sophisticated.
Sophistication is not evidence.
Every extra layer should explain an independent observation.
And every extra free parameter increases the chance of fitting by accident.
A good layered model therefore needs held-out tests.
Fit one set of pages.
Freeze the rules.
Predict another set.
The Voynich problem has enough flexibility already.
A successful theory must reduce freedom, not add an elegant name to it.
What Would Falsify the Directional Dissociation?
A strong result should be willing to disappear.
- If the direction effect vanishes under a better glyph segmentation, it was partly a transcription artefact.
- If it appears equally in broad historical controls once the same preprocessing is used, it is less diagnostic.
- If it is driven by one quire or one Currier regime, it is not manuscript-wide.
- If uncertain spaces account for most cross-boundary dependence, the boundary layer needs reinterpretation.
- If a simpler generator reproduces the full joint signature under independent implementation, the cipher-like interpretation weakens.
Falsifiability is not hostility to the result.
It is what turns an interesting curve into useful knowledge.
Primary School: Two Rules in One Game
Imagine a game where letters inside each box must follow one rule, but the boxes themselves must follow another.
For example:
- inside each box, letters must end in ABC order;
- between boxes, colours must alternate red-blue-red-blue.
Looking only inside the boxes reveals one system.
Looking only between boxes reveals another.
That is the directional dissociation idea in miniature.
Secondary School: Reverse the Model, Not the Manuscript
Take a set of artificial tokens.
Measure character predictability left-to-right.
Then reverse each token and measure again.
Do the same for the sequence of token boundaries.
Students learn that “statistically easier backwards” is not the same statement as “historically read backwards”.
JC and Adult Readers: Build a Two-Layer Direction Test
Keep two measures separate.
- Internal directionality: predictive asymmetry among characters within the same token.
- Boundary directionality: predictive dependence from the end of one token into the beginning of the next.
Then test both under:
- alternate transcriptions;
- Currier A/B;
- page-type controls;
- space-confidence thresholds;
- synthetic cipher and generator controls.
A model that explains only one layer should not be called a manuscript-scale mechanism.
Reader Checklist: Before You Say Voynich “Runs Backwards”
- Are you discussing physical writing direction or statistical directionality?
- Is the result inside tokens, across boundaries, or both?
- Which transcription was used?
- How were compound glyphs treated?
- How were uncertain spaces treated?
- Does the effect survive Currier A/B separation?
- Does it survive quire-level resampling?
- Which comparison languages were used?
- Which synthetic generators were tested?
- Did those generators fail one statistic or a joint criterion?
- Is “cipher-like” being promoted to “cipher proven”?
- What observation would make the directional result disappear?
Frequently Asked Questions
Does Voynich read right-to-left?
The physical prose is overwhelmingly written left-to-right. The 2026 result concerns statistical predictability inside visible token sequences, not a demonstrated historical reading direction.
What is directional dissociation?
It is the reported split in which token-internal character structure shows one directional optimization while cross-token boundary dependence shows the opposite directional tendency.
Does this prove Voynich is a cipher?
No. The result is described as suggestive of cipher-like layered constraints, but other transformed or multi-stage systems may also produce such architecture.
Has this been peer reviewed?
The Parisel work discussed here is a 2026 arXiv preprint. It should be treated as a current quantitative result awaiting wider independent replication and review.
Why is it useful anyway?
Because it supplies a specific joint behaviour that competing mechanisms can try to reproduce. Even if later refined, it raises the bar from matching one surface statistic to explaining multiple structural layers at once.
Research Foundations
- Christophe Parisel — Evidence of Layered Positional and Directional Constraints in the Voynich Manuscript (2026 preprint).
- Liudmila Rozanova & Alexander Temerev — A Glyph Is Not a Letter, a Token Is Not a Word, a Space Is Not a Space (2026 preprint).
- René Zandbergen — The Voynich Writing System.
- Which Way Does Voynich Read?.
The Final Idea
The Voynich Manuscript may be giving us two arrows at once.
One belongs to the internal construction of its token-like units.
One belongs to what happens when those units meet.
That does not tell us the language.
It does something more useful at this stage.
It tells every future explanation that one direction and one layer may no longer be enough.