Voynich research inherited two units before it had earned either one.
The glyph.
The word.
We see a shape, so we name a character.
We see a blank, so we name a word boundary.
Those choices are reasonable.
They may also be too coarse in one direction and too fine in the other.
In August 2026, Rozanova and Temerev reported a result that makes this question unavoidable. In their analysis, Voynich’s unusually low character-level uncertainty did not behave like ordinary one-letter substitution. Instead, recurring structure resolved onto a stable scale of multi-symbol units—larger than one glyph, smaller than the visible space-delimited token, and stable enough to recur across quires.
The paper is a preprint.
The result needs replication.
But the research question is fundamental even if the exact unit inventory changes.
What if Voynich’s most natural building blocks are not the units our transcription gives us, but structures the manuscript repeatedly assembles across those units?
Quick Read
- A 2026 preprint reports that Voynich’s character-level regularity is too strong for straightforward one-to-one substitution of the tested plaintext controls.
- The same study reports a quire-stable scale of recurrent multi-symbol units.
- These units are larger than individual transliterated glyphs but can be smaller than the visible token between spaces.
- This does not mean the authors have discovered syllables, morphemes or plaintext letters.
- A multi-symbol unit can arise from ligature structure, abbreviation, verbose cipher groups, recurring grapheme clusters, generated chunks or true linguistic subunits.
- The result is especially important because some learned units cross uncertain spaces after all spaces have been erased before learning.
- That links the unit-scale problem directly to the Edge Problem and to graded boundary strength.
- If recurrent units remain stable across quires, they may be more fundamental than many exact whole-token forms.
- A successful mechanism should explain why these units recur, how they combine, where they are permitted and whether their inventory changes by Currier regime or hand.
- The correct conclusion is not “Voynich units are syllables.” It is “glyph and word may not be the only scales that deserve explicit modelling.”
Why Unit Choice Matters More Than It Looks
Every statistic depends on the unit being counted.
Count characters and you get character entropy.
Count words and you get word frequencies.
Count morphemes and you get another system entirely.
If the wrong unit is chosen, real structure can look bizarre.
Imagine English written without spaces:
therainfalls
If we arbitrarily cut it as:
ther / ain / falls
our “words” become strange even though the underlying system is ordinary.
Voynich gives us the reverse problem too.
We may have cut correctly at many blanks while still failing to identify stable subunits inside the tokens.
Glyph Is a Visual Unit, Not Automatically a Functional Unit
The existing Alphabet Problem already shows why one visible form does not automatically equal one linguistic character.
A shape can be:
- one grapheme;
- an allograph;
- a ligature;
- an abbreviation;
- a compound;
- a modifier plus base;
- a transcription convenience.
The Unit-Scale Problem asks the complementary question.
Even if our glyph segmentation is visually useful, do several consecutive glyphs repeatedly behave as one higher-order structural object?
What Is a Recurrent Multi-Symbol Unit?
Suppose a sequence such as ABC appears again and again inside many larger tokens.
Suppose AB almost never appears without C.
Suppose ABC occurs in several word families and across several quires.
Then ABC may be a better statistical unit than A, B and C considered independently.
It still may not be one linguistic unit.
It may be:
- a syllable-like cluster;
- a morpheme-like cluster;
- a verbose cipher group;
- a scribal abbreviation unit;
- a recurring graphical chunk;
- a generator slot;
- a ligature represented as several transliteration symbols.
The structural observation comes first.
The linguistic name comes later, if earned.
Why Quire Stability Is Important
A unit discovered only on one page may be a local accident.
A unit discovered across many quires is harder to dismiss.
The 2026 preprint reports that the characteristic recurrent-unit scale is stable at quire level under its resampling design.
This matters because Voynich varies strongly by section, hand and Currier regime.
If one structural scale survives those changes, it may belong to a deeper writing-system constraint.
That does not mean every quire uses the same unit frequencies.
It means the scale at which repeated chunks become informative may remain surprisingly stable.
Why One-to-One Substitution Struggles
A simple substitution cipher replaces each plaintext character with one ciphertext character.
Such substitution preserves much of the source language’s character-level structure.
Rozanova and Temerev report Voynich conditional entropy around 2.7 bits in their setup, lower than their Latin, Italian and English plaintext controls near 3.5 bits.
They interpret that as too constrained for straightforward one-to-one substitution of those tested plaintexts.
The result does not eliminate ciphertext.
It changes the kind of ciphertext mechanism that remains plausible.
- verbose substitution;
- multi-symbol code groups;
- homophonic expansion;
- abbreviation plus cipher;
- stateful encoding.
All of those can move functional information away from one-glyph-per-letter assumptions.
The Bench Characters Are an Obvious Unit-Scale Test
Voynich benches conventionally transcribed with pairs such as ch and sh already expose the problem visually.
Are they two characters?
One ligature?
A base plus modifier?
If the statistical learner repeatedly treats a bench sequence as one recurrent unit, that supports—but does not prove—a fused functional interpretation.
The Bench Characters article owns the palaeographic problem.
Unit-scale analysis supplies an independent distributional test.
The Minim Strings Are Another Test
Sequences conventionally represented as i, ii, iii and related forms are visually repetitive.
Do they represent repeated independent signs?
One expandable unit?
Counting notation?
A scribal compression?
A learned multi-symbol model can ask whether these strokes travel together often enough to deserve a higher-level representation.
Again, statistics cannot name the unit.
It can tell us whether splitting it destroys regularity.
The Q-Series May Be a Multi-Symbol Unit Too
Voynich q is famous because it strongly prefers o-like following material.
That raises a familiar question.
Is q one sign whose syntax demands o?
Or is qo one functional unit that transcription happens to split?
The Q-Series article owns the positional facts.
The Unit-Scale Problem adds a representation test:
does a model become simpler and more stable when qo-like clusters are treated as higher-order units?
Units That Cross Uncertain Spaces
This is one of the most important 2026 observations.
The Edge Problem reports that spaces marked uncertain by transcribers are physically narrower and behave differently from confident spaces in the preprint’s tests.
Rozanova and Temerev also erase every space before learning recurrent units.
Some learned units then cross locations where the transcription had inserted an uncertain space.
This is exactly what we would expect if at least some uncertain gaps split a larger functional unit.
It is not what we would expect if every uncertain gap were simply a noisy version of a true word boundary.
See The Edge Problem.
Could the Units Be Syllables?
Possible.
Not established.
Syllabic scripts naturally use recurring multi-symbol or multi-stroke structures depending on transcription.
But without sound values, calling a recurrent chunk a syllable merely translates “multi-symbol unit” into a linguistic word.
A syllable hypothesis needs independent phonological evidence.
The Vowel Problem shows why that evidence is currently weak.
Could the Units Be Morphemes?
Also possible.
Voynich word families often look like a stable core surrounded by additions or substitutions.
A morpheme-like account would predict recurring chunks associated with stable positional or grammatical roles.
But semantic morphology requires more than recurrence.
The same unit should combine with different families according to a repeatable rule.
Its distribution should predict something downstream.
Until then, “morpheme-like” remains a useful hypothesis class.
Could the Units Be Cipher Groups?
This is one of the strongest non-linguistic interpretations.
A verbose cipher can map one plaintext unit into several visible symbols.
Those symbols can have restricted internal order.
Whole visible tokens can then be assembled from several code groups.
Under such a mechanism, multi-symbol units are not syllables or morphemes.
They are encoding blocks.
The new directional and edge results make this especially interesting because a code group could obey one internal rule while boundaries between groups obey another.
Could the Units Be Generator Slots?
A pseudo-text generator can also create recurring chunks.
A slot system may repeatedly choose from a small inventory of prefixes, middles and endings.
Self-citation can propagate multi-symbol fragments from one token to descendants.
Therefore unit-scale evidence cannot distinguish meaning from non-meaning by itself.
It tells us what the generator—whatever it is—reuses.
The historical mechanism remains the next question.
A Better Representation Could Simplify Voynich
Imagine replacing a repeated three-glyph sequence with one unit label.
The visible alphabet becomes smaller at one scale and the token grammar changes.
Entropy changes.
Word length changes.
Boundary statistics change.
Some apparently rare tokens become common combinations of common units.
This is why the correct unit representation can turn a mysterious corpus into a simpler one.
It is also dangerous.
If researchers invent units after seeing which ones improve their preferred model, they can overfit.
Unit learning must therefore be algorithmic, blinded or held out where possible.
The Unit Inventory Must Generalise
A useful unit set should not work only on the pages used to discover it.
Learn units on one collection of quires.
Freeze the inventory.
Apply it to held-out quires.
Then ask:
- Do the same units recur?
- Do they explain comparable amounts of structure?
- Do new Currier regimes require a different inventory?
- Do labels use the same units as prose?
- Do rare glyphs create special units?
Generalisability is what separates a discovered structure from a fitted convenience.
The Unit-Scale Problem and the Hapax Problem Fit Together
Voynich can have many singleton complete tokens while using a small inventory of recurrent subtoken units.
That would resolve part of the paradox.
Surface vocabulary stays open.
Underlying construction stays constrained.
This is not proof of any one mechanism.
It is a clear architectural possibility that every proposed mechanism should now confront.
See The Hapax Problem.
What the Unit-Scale Problem Does Not Prove
- It does not prove the units are syllables.
- It does not prove they are morphemes.
- It does not prove a cipher.
- It does not prove a generator.
- It does not prove all visible glyphs are mis-segmented.
- It does not prove all visible spaces are wrong.
- It does not make one unit-learning algorithm ground truth.
What It Can Give Us
- A candidate structural level between glyph and token.
- A way to test whether recurring chunks generalise across quires.
- A bridge between palaeography, entropy, tokenisation and cipher models.
- A way to reduce vocabulary complexity without pretending to translate it.
- A new falsification target for every mechanism family.
Reader Checklist
- How were the units learned?
- Were spaces removed first?
- Which transcription was used?
- Are units stable across quires?
- Do they survive held-out testing?
- Do they cross uncertain spaces more than confident spaces?
- Do they correspond to visible ligatures or benches?
- Does the inventory change by Currier regime?
- Does a proposed cipher or generator predict the same units?
- Is a statistical chunk being prematurely named a syllable or morpheme?
Research Foundations
- Rozanova & Temerev — A Glyph Is Not a Letter, a Token Is Not a Word, a Space Is Not a Space (2026 preprint).
- The Edge Problem.
- The Alphabet Problem.
- The Bench Characters.
The Final Idea
The Voynich Manuscript may be difficult partly because we keep asking it to speak in units convenient for us.
One glyph.
One word.
One space.
The manuscript may be organised at another scale.
Before translating Voynich units, we may still have to discover which units the manuscript itself keeps choosing to reuse.