Voynich Sentence Boundaries: Punctuation, Paragraphs and the Missing Full Stop

Where does a Voynich sentence end?

It is one of those questions that seems as though it should have been settled before anyone began discussing language.

Find the punctuation.

Find the sentence.

Count the words.

Analyse the grammar.

Except the Voynich Manuscript never grants us that sequence so politely.

Paragraphs are real enough.

Most ordinary running text is arranged top-to-bottom and left-to-right in short paragraphs.

Paragraphs are often separated by larger vertical gaps.

Their first characters can be enlarged or decorated.

Their final lines are usually shorter than the preceding lines.

Sometimes the last line is centred.

Sometimes right-justified.

Sometimes a short additional element sits at the right end of the final line.

These are visible document structures.

Sentences are not.

There is no accepted full-stop-like sign that reliably tells researchers:

the proposition ends here; the next proposition starts there.

This matters far beyond punctuation.

Syntax depends on unit boundaries.

Phrase repetition depends on unit boundaries.

Language modelling depends on unit boundaries.

Even the meaning of “word order” changes if one line contains several sentences, one sentence spans several lines, or the text is not sentence-based prose at all.

So the proper question is not:

Why did the Voynich writer forget punctuation?

It is:

What boundary system does the manuscript actually expose, and at what level does each boundary operate?

That keeps modern grammar from being smuggled into an undeciphered script.

Quick Read

One-sentence answer: the Voynich Manuscript has clear line and paragraph structure but no accepted sentence-delimiting punctuation in its principal script; paragraph boundaries are supported by layout, enlarged initials, spacing and characteristic final lines, while any sentence-level segmentation remains a hypothesis that must be inferred from recurring textual behaviour rather than imposed from modern punctuation expectations.

Most ordinary Voynich prose runs left-to-right and top-to-bottom.
Running text is organised into short paragraphs separated by larger line spacing.
Paragraph-initial characters come from a restricted subset and are often enlarged or embellished.
Paragraph-final lines are commonly shorter than preceding lines.
Some paragraph ends use centred lines.
Some use right-justified lines.
Some contain a left-justified ending plus additional right-justified words, a layout John Grove called “titles”.
These layouts identify paragraph-final behaviour, not decoded punctuation.
No accepted principal-script sign functions securely as a universal modern-style full stop.
No accepted principal-script sign securely marks commas, semicolons, questions or sentence boundaries manuscript-wide.
Visible spaces appear to divide word-like groups, but those groups are not proven lexical words.
Line breaks are strong physical units but are not proven sentence boundaries.
Paragraph breaks are stronger document boundaries than any proposed sentence break.
Labels, circular text, radial text and star-marked entries use different document structures and should not inherit prose punctuation assumptions.
Medieval punctuation practices were variable, so the absence of modern punctuation is not itself anomalous enough to identify the script.
A real sentence model should predict recurring boundary-sensitive grammar or token classes across text not used to invent the segmentation.

For now, the manuscript gives us paragraphs more confidently than sentences.

That asymmetry is evidence.

A Paragraph Is Visible Because Several Signals Agree

Paragraph recognition does not depend on translating anything.

Several independent visual cues converge.

larger spacing before the next block;
new line beginning at the left margin;
restricted or enlarged initial glyphs;
shortened final line;
occasional special final-line alignment.

When these signals agree, paragraph status is not merely a transcriber convenience.

It belongs to the physical manuscript.

This is why modern IVTFF transcription explicitly preserves paragraph starts and distinguishes several types of final-line layout.

The transcription is not inventing the paragraph.

It is encoding a visible document boundary.

That makes paragraph a strong analysis unit even before sentence is solved.

The Paragraph Initial Behaves Like a Boundary Signal

Paragraph beginnings are not statistically ordinary positions.

Only a restricted subset of Voynich characters tends to occur at paragraph starts.

Some initial forms are larger.

Some have additional curls or dots.

This is compatible with familiar manuscript practices such as initials or display forms.

It may also carry structural information within Voynichese itself.

The important point is that the boundary is double-coded.

Layout changes.

Character distribution changes.

That is much stronger evidence for a document boundary than a rare dot-like mark noticed once beside one token.

A real sentence delimiter should eventually achieve something comparable:

its presence should predict a reproducible change in what text can appear before and after it.

Paragraph Final Lines Are Not Ordinary Lines Either

Most paragraph-final lines are shorter.

That can happen naturally because the paragraph simply runs out of text.

But some Voynich paragraphs use more conspicuous closing formats.

René Zandbergen’s survey, following John Grove’s terminology, records pages with:

centred final lines;
right-justified final lines;
final lines containing a separate right-justified element.

The exact interpretation remains open.

Closing formula?

Heading?

Tag?

Record identifier?

Visual flourish?

What survives is the boundary-sensitive placement.

The final line is doing something different from the middle of the paragraph.

That makes paragraph end a natural place to search for sentence- or discourse-closing grammar.

A Line Is Not Automatically a Sentence

This is perhaps the most important negative rule.

Voynich lines are physically obvious.

They also have strong statistical properties.

Line beginnings favour some glyphs and token families.

Line endings favour others.

But that does not mean each line expresses one sentence.

Image geometry often determines available line width.

A plant branch can shorten a line.

A diagram can force a new line.

Foldout geometry can change text flow.

If line end were automatically sentence end, picture geometry would repeatedly decide where propositions terminate.

That is possible in a specialised notation.

It is not a safe default for language.

The line must therefore remain a production/layout unit until sentence equivalence is demonstrated.

A Paragraph Is Not Automatically a Sentence Either

Paragraphs are stronger boundaries than lines.

They still need not correspond one-to-one with modern sentences.

A paragraph can contain:

several sentences;
one long sentence;
a recipe entry;
a list of fields;
a catalogue record;
a formulaic note;
another non-sentence information unit.

This matters especially because parts of Voynich may not be prose in the ordinary narrative sense.

Quire 20 may contain record-like entries.

Labels are not sentences.

Circular text may operate as annotations around sectors.

A sentence model should therefore be restricted to text roles where sentence grammar is plausible.

Document role comes before universal punctuation.

Why We Should Not Search for a Modern Full Stop by Shape Alone

One tempting strategy is visual.

Find dots.

Find small circles.

Find rare marks between tokens.

Declare punctuation.

The problem is that the manuscript contains:

rare glyphs;
decorative dots;
paragraph-initial embellishment;
later marginal marks;
damage;
paint;
star-like record markers.

Shape alone cannot establish syntactic function.

A punctuation candidate should have a distribution.

It should appear between text units rather than inside ordinary word families.

What follows should differ statistically from ordinary continuation.

Its frequency should be plausible for the unit it supposedly marks.

A sentence delimiter should delimit sentences.

Visual resemblance to punctuation is only the beginning.

Medieval Punctuation Does Not Need to Look Modern

The missing modern full stop is less surprising once the historical comparison is widened.

Medieval manuscript punctuation was variable across languages, periods, institutions and individual scribes.

Different point systems could mark different degrees of pause.

Visual paragraphing could perform work that modern punctuation performs with marks.

Capital or enlarged initials could help orient readers.

Specialised technical genres might rely more on layout than prose punctuation.

Therefore:

absence of a modern punctuation system is not evidence that the text cannot be medieval language.

At the same time, the historical comparison does not give Voynich a free pass.

If it encodes ordinary connected language, grammatical boundaries still have to exist somewhere, whether visibly punctuated or not.

Sentence Boundaries Can Be Invisible in Writing

Not every language or writing tradition needs explicit sentence punctuation.

Readers can recover boundaries from grammar and discourse.

A sentence may end because:

the verb construction closes;
a new subject structure begins;
a discourse particle opens a new clause;
a formulaic ending recurs;
the semantic unit is complete.

This creates a route for Voynich research.

Do not find punctuation first.

Find recurring boundary grammar.

If one token family appears strongly before paragraph ends, another strongly after them, and the same pair also occurs internally at plausible clause transitions, sentence-like segmentation may emerge distributionally.

The syntax can reveal the boundary even if the scribe did not mark it with a point.

Paragraph Ends Are the Best Calibration Set

We do know where many paragraphs end.

That gives us labelled boundary data before language is decoded.

Compare tokens in three positions:

ordinary line-internal positions;
ordinary line ends inside paragraphs;
true paragraph-final positions.

If a form appears only at paragraph ends, it may be a discourse closer, paragraph marker or special orthographic ending.

If it also appears at internal line ends, it may instead be margin-sensitive.

If it appears internally before a recurring new-start class, it becomes a candidate sentence or clause closer.

This is a much stronger path than searching visually for dots.

Known paragraph boundaries can calibrate unknown sentence boundaries.

Centred Final Lines May Be Closers, but We Do Not Know Their Grammar

Centred paragraph-final lines are visually conspicuous.

They invite functional stories.

Perhaps they are headings.

Perhaps conclusions.

Perhaps recipe names.

Perhaps signatures or tags.

None is established.

The safer route is comparative.

Ask whether centred final lines have:

different vocabulary;
different token lengths;
different initial or final character families;
different Currier distribution;
different recurrence across pages.

Layout first.

Textual register second.

Meaning last.

The same discipline used for labels should apply here.

Right-Justified Final Lines May Be Another Register

Right-justified paragraph endings are unusual enough to deserve separate analysis.

They are not merely short lines.

The scribe intentionally places them away from the usual left margin.

That means layout is carrying a category signal.

Again, possible functions are numerous.

closing formula;
heading for what precedes;
heading for what follows;
identifier;
annotation;
visual balance.

A robust theory should not assign function from position alone.

It should show that the same textual family behaves consistently whenever that position is used.

Special alignment is a structural label supplied by the page.

The text still has to tell us what the label means.

John Grove’s “Titles” Are Especially Important

John Grove used the term “titles” for a distinctive configuration in which a normal left-justified final line has additional material at the right.

The name is descriptive shorthand.

It should not be mistaken for decoded function.

What makes these cases interesting is their distribution.

Centred and right-justified final lines occur across the main handwriting styles.

Strict Grove-style titles, by contrast, have been reported only on Currier-A / hand-1 pages in Zandbergen’s survey.

That correlation may connect document layout to textual regime or production practice.

It does not tell us what the “title” says.

But it is exactly the kind of cross-axis relationship that can reveal hidden document grammar.

Could Gallows Mark Sentence Starts?

Gallows characters are strongly associated with special positions, especially paragraph and line beginnings.

This naturally invites a capital-letter analogy.

Perhaps gallows begin sentences.

The hypothesis is testable.

If gallows mark sentence beginnings, internal gallows occurrences should align with independent evidence of a new clause or sentence.

The same relation should survive different line widths and document roles.

But gallows are also tied strongly to physical line/paragraph position.

That means a layout explanation competes directly with a sentence-start explanation.

Until these causes are separated, gallows cannot serve as sentence capitals by assumption.

Position is evidence.

The type of position is the unresolved question.

Could One Common Ending Be a Full Stop in Disguise?

Another possibility is that punctuation is integrated into word-like forms.

A token-ending sign may indicate sentence closure.

A special suffix may function as a discourse final.

A terminal allograph may combine orthography and punctuation.

This would explain why no standalone full-stop sign is obvious.

The model has a strong prediction.

The candidate ending should be enriched before known paragraph boundaries and before independently inferred internal sentence boundaries.

It should not simply occur at every physical line end.

This could connect the minim-ending system or other final families to discourse structure.

For now, no such mapping is accepted.

But sentence punctuation may be morphological rather than separate.

Could Spaces Carry More Than One Boundary Strength?

Voynich spaces are already under investigation as possible word boundaries.

Not all visible gaps need have the same function.

One gap could separate subword groups.

A larger gap could separate words.

An even larger gap might separate clauses.

This is a plausible hierarchical boundary model.

It requires geometry rather than ordinary transcription alone.

Exact gap width has to be measured from images.

Then surrounding token classes should be tested.

A digital text file that turns every gap into one identical ASCII space destroys this possibility.

The sentence problem may therefore live partly in layout, not only characters.

Labels Should Not Be Forced Into Sentence Grammar

Labels form a specialised textual population.

Most are short.

Many are unique.

They occupy visual positions rather than paragraph flow.

Asking where a label’s “sentence” ends is probably the wrong question.

The unit may be an identifier.

Coordinate.

Name.

Property.

Likewise a radial annotation may be one field rather than one sentence.

Document grammar should precede sentence grammar.

Otherwise prose expectations contaminate every text role.

Quire 20 May Use Entry Boundaries Instead of Sentences

The star-marked pages sharpen the same issue.

Marginal star-like markers repeatedly align with short entry-like blocks.

The star can therefore operate as a record boundary or address marker.

Inside each record, ordinary sentence punctuation may be unnecessary if the text is formulaic.

A modern database row does not need a period after every field.

A recipe entry may use a sequence of conventional fields.

A registry can rely on repeated layout.

This means “missing punctuation” may partly be a genre mismatch.

If Quire 20 is record-like rather than prose-like, entry segmentation is the first boundary to recover.

Circular Text Has No Natural Full Stop at the Right Margin

Circular text makes sentence assumptions even more fragile.

A circle has no right margin.

There is no physical line end unless the text itself marks one.

Modern transcription has to choose a cut point for the circle.

That cut can accidentally masquerade as a boundary.

The correct analysis should preserve wrap-around adjacency.

If the last transcribed token and first transcribed token are physically neighbours on the ring, they must remain neighbours in the model.

Sentence segmentation cannot be allowed to inherit the arbitrary point where a modern editor started typing.

Reading direction and rotation therefore belong upstream of punctuation in circular loci.

Which Way Does Voynich Read? owns the direction, rotation and wrap-around geometry that must be resolved before circular text can be segmented responsibly.

Sentence Segmentation Could Strengthen Syntax

Reddy and Knight found relatively weak exact whole-word bigram predictability in a Currier-B sample compared with several natural-language controls.

But if modern token sequences cross unmarked sentence boundaries freely, some bigrams are being measured across places where grammar resets.

This effect exists in every language corpus if punctuation is removed.

Recovering sentence boundaries could therefore sharpen syntactic dependencies.

Or it might not.

If no plausible segmentation produces stronger grammar, sentence-style prose becomes less attractive as a model.

This makes sentence segmentation testable through predictive improvement.

A proposed boundary system should make what comes next easier to predict under a coherent model.

Sentence Segmentation Could Also Create False Grammar

The opposite danger is equally serious.

If researchers are free to insert sentence boundaries wherever grammar looks weak, the segmentation will always improve the model.

Break before every surprising token.

Call the surprise a new sentence.

Soon the remaining sequences look orderly.

That is circular fitting.

A serious boundary model should be learned from one set of visible and statistical cues, then applied to unseen passages.

The rule must choose boundaries before the downstream grammar score is known.

Otherwise “sentence” becomes a repair tool rather than a discovered unit.

A Language Model Makes Clear Predictions About Boundaries

If Voynich encodes natural language, sentence boundaries should reveal recurring grammar.

Possible signals include:

special sentence-initial function words;
restricted opening morphology;
closing particles;
verb-final or verb-initial patterns;
discourse markers;
capital-like display forms.

We do not need to know their meanings first.

Latent classes can be measured.

If one class repeatedly appears after inferred boundaries and another before them, the language hypothesis gains structural support.

The same grammar should work in multiple sections after Currier and document role are controlled.

Sentence structure should generalise.

A Cipher Model Inherits the Plaintext Boundary Problem

Ciphertext can obscure punctuation.

A cipher may encode punctuation as special symbols.

Ignore it.

Absorb it into code groups.

Or remove it before encryption.

Therefore lack of visible punctuation is compatible with cipher.

But a complete cipher solution should restore source-side sentence structure after decryption.

If the plaintext remains an endless undifferentiated word stream, the cipher has not solved the discourse problem.

A strong cipher should explain why paragraph layout survives even if sentence punctuation does not.

Again, the absence creates a burden rather than a disproof.

A Generated-System Model May Not Need Sentences at All

A templatic or procedural generator can produce paragraphs without sentence grammar.

Create a paragraph start.

Generate legal tokens under local constraints.

Terminate when space or another condition is reached.

No sentence delimiter is required.

This naturally explains paragraph-level structure with weak sentence evidence.

But the generator must then explain:

why paragraph starts are special;
why paragraph endings are special;
why centred and right-justified ending registers exist;
why Currier and document-role structure persist;
why long-range vocabulary organisation survives.

No-sentence grammar is compatible with generation.

It is not an explanation of the rest of the manuscript by itself.

A Record System May Use Fields Instead of Sentences

This is one of the strongest middle-ground possibilities.

Not ordinary prose.

Not meaningless text.

Structured records.

A record can contain fields in a fixed order.

Ingredient.

Quantity.

Operation.

Condition.

No period is needed between every field.

Paragraph and entry boundaries may matter more than sentence boundaries.

This model is especially relevant to Quire 20 and other formulaic-looking text roles.

It also predicts stronger class-level order than exact phrase repetition—something already observed as a promising scale.

Document grammar may be the hidden punctuation system.

What Survives the Sentence-Boundary Work

Voynich ordinary running text has strong visible paragraph structure.
Paragraph boundaries are often supported by spacing, special initials and shortened final lines.
Centred, right-justified and Grove-style final-line layouts form reproducible paragraph-end phenomena.
Lines are physically real and statistically important but are not proven sentences.
Paragraphs are stronger units than sentences but need not correspond one-to-one with sentences.
No accepted principal-script punctuation system securely marks sentence boundaries manuscript-wide.
Visual dot-like or rare marks cannot be declared punctuation from resemblance alone.
Sentence boundaries may be implicit in grammar, integrated into token endings or reflected in spacing rather than separate symbols.
Labels and radial/circular text should not inherit prose sentence assumptions.
Quire 20 may use entry or field boundaries more naturally than sentence boundaries.
A valid sentence model should improve predictive grammar on unseen text without moving boundaries after the fact.

What Does Not Survive as Established Knowledge

Each Voynich line is one sentence.
Each Voynich paragraph is one sentence.
The manuscript has no punctuation of any kind.
A specific rare glyph is proven to be a full stop.
Gallows are proven sentence capitals.
Every centred final line is a title.
Every right-justified final line is a closing formula.
Visible spaces are proven lexical word boundaries.
Sentence grammar must apply uniformly to labels and records.
Lack of modern punctuation proves the text is not natural language.

The document boundaries survive.

The sentence map does not.

A Better Sentence-Boundary Analysis

Start with visually secure paragraph boundaries.
Separate ordinary line ends from paragraph ends.
Tag centred, right-justified and Grove-style final lines separately.
Measure paragraph-initial and paragraph-final token classes.
Do not declare punctuation from shape alone.
Measure gap width if hierarchical spacing is proposed.
Control page geometry and line width.
Separate prose, labels, circular text, radial text and records.
Infer candidate sentence boundaries from reproducible before/after distribution changes.
Freeze the boundary rule before testing downstream syntax.
Require the segmentation to improve prediction on unseen pages.
Allow the possibility that some Voynich registers are not sentence-based at all.

What Would Count as a Real Sentence-Boundary Breakthrough?

Imagine researchers build a boundary model using only known paragraph ends and local token distributions.

The model discovers two recurrent classes.

One class concentrates immediately before paragraph boundaries.

Another concentrates immediately after them.

The same paired transition occurs repeatedly inside long paragraphs.

A boundary hypothesis is frozen.

On unseen pages, those internal boundaries predict stronger class-level syntax and reduce impossible transitions.

A later decipherment independently reveals clause or sentence boundaries in the same places.

That would be a genuine sentence breakthrough.

No full stop would have needed to be found.

The boundary would have been recovered from the system’s behaviour.

The strongest punctuation mark may turn out not to be a mark at all, but a reproducible change in what the writing system allows next.

Primary School: A New Paragraph Is Not Always a New Sentence

Give a child a short story with punctuation removed.

Keep paragraph breaks.

Ask where the sentences probably end.

The paragraph helps.

It does not tell us every sentence boundary.

This is the Voynich problem in miniature.

Lower Secondary: Separate Boundary Levels

Give students one page containing:

word spaces;
line breaks;
paragraph breaks;
sentence punctuation.

Remove one level at a time.

Ask what information is lost.

Students discover that boundaries form a hierarchy rather than one universal delimiter.

Voynich gives us some levels more clearly than others.

Upper Secondary: Infer an Invisible Boundary

Create an artificial text where a special class of word almost always begins a sentence but punctuation has been removed.

Let students discover the distribution.

Then test the rule on new text.

The exercise demonstrates how syntax can recover structure that visual punctuation does not mark.

JC and Adult Readers: Boundary Recovery as Latent Segmentation

At a higher level, sentence segmentation can be treated as a latent-variable problem.

Observed variables include:

token classes;
line position;
gap size;
paragraph position;
special alignment;
Currier regime;
document role.

The hidden variable is boundary state.

A successful model should discover a small number of states that improve sequence prediction and recur across independent text.

Crucially, the number of boundaries should be penalised.

Otherwise the model can insert a boundary before every difficult transition and fit perfectly.

The problem is therefore not merely segmentation.

It is segmentation under complexity control.

A Parent and Teacher Guide

Separate word, line, paragraph and sentence boundaries.
Use visible paragraphs as anchors.
Do not call every line a sentence.
Do not search only for dot-shaped punctuation.
Remember that medieval punctuation need not look modern.
Keep labels and records outside prose grammar until justified.
Make a boundary rule predict what comes before and after.
Freeze the rule before testing syntax.

The transferable lesson is simple:

when punctuation is missing, do not invent sentences from intuition; recover boundaries from the structure that changes across them.

Reader Checklist: Before You Segment Voynich Into Sentences

Is the text ordinary paragraph prose, label text, circular text or record-like text?
Where are the secure paragraph boundaries?
Which lines are paragraph-final?
Is the line shortened by image geometry?
Is the candidate delimiter an ordinary glyph, rare glyph or later mark?
Does the candidate occur consistently between units?
Does what follows differ statistically from ordinary continuation?
Do centred and right-justified endings form separate registers?
Could sentence closure be integrated into token endings?
Could different gap widths mark different boundary strengths?
Does the boundary rule survive Currier A and B?
Does it survive different hands?
Does it improve prediction on unseen pages?
Could the text role use fields rather than sentences?

Frequently Asked Questions

Does the Voynich Manuscript have punctuation?

It has clear layout boundaries and special positional forms, but no accepted principal-script punctuation system has been shown to mark modern-style sentence boundaries consistently.

Are Voynich line breaks sentence breaks?

Not established. Lines are strong physical and statistical units, but page geometry often determines where they end.

Are paragraphs easy to identify?

Usually much easier than sentences. Larger line spacing, paragraph initials and characteristic shortened final lines provide strong visual signals.

What are the centred and right-justified final lines?

They are reproducible paragraph-final layouts. Their exact function—closing formula, title, identifier or something else—has not been decoded.

Could gallows be capital letters at sentence starts?

It is a possible analogy, but gallows are strongly linked to line and paragraph position. Sentence-start function has not been demonstrated independently of layout.

Could sentence boundaries be hidden inside words?

Possibly. A special ending, allograph or discourse-closing token could mark boundaries without separate punctuation. Such a model would need strong boundary-specific distributions.

Why does this matter for decipherment?

Because syntax, phrase repetition and grammar depend on knowing where units begin and end. Wrong sentence segmentation can either hide real structure or manufacture false structure.

What is the strongest current conclusion?

Voynich paragraphs are real document units; sentence boundaries are not yet securely identified. A serious solution should recover sentence or field boundaries from repeatable system behaviour rather than impose modern punctuation conventions.

Research and Further Reading

The Final Idea

The Voynich Manuscript does not leave us without boundaries.

It gives us many.

Spaces.

Lines.

Paragraphs.

Labels.

Star-marked entries.

Circular sectors.

What it does not give us securely is the modern sentence.

That may be because punctuation is implicit.

Because punctuation is encoded.

Because sentence grammar is unusual.

Or because some parts of the manuscript are not sentence-based documents in the first place.

The temptation is to insert our own full stops so analysis can begin.

The better discipline is the reverse.

Let the manuscript earn its sentence boundaries by showing us that the system behaves differently on either side of them.

Continue Through the Voynich Research Map

This article is one specialist node in eduKateSG’s larger Voynich research library. Return to the canonical master for the manuscript-wide evidence architecture, or continue into the line, space and direction problems that constrain sentence recovery.

Paragraphs are visible document units; sentence boundaries must be earned from repeatable system behaviour rather than imposed from modern punctuation expectations.

Voynich | Everything eduKate Knows and Tested | Where Does a Sentence End? Punctuation, Paragraphs and the Missing Full Stop