Voynich Semantic Grounding: How Does Structure Become Meaning?

Voynich research is getting better at finding structure.

That is good.

We can measure token families.

Character positions.

Currier regimes.

Boundary effects.

Local vocabulary.

Visual clusters.

Labels.

Possible homophone-like equivalence classes.

Possible cross-references.

Possible notation states.

Then comes the dangerous step.

Meaning.

A statistical cluster becomes “plant”.

A common token becomes “the”.

A label beside a star becomes the star’s name.

A repeated ending becomes a grammatical case.

A vessel becomes “distillation”.

A boundary marker becomes “and”.

Sometimes one of those assignments may eventually prove correct.

But the assignment itself is not evidence merely because the underlying structure is real.

The Semantic-Grounding Problem is the problem of crossing from “this pattern exists” to “this pattern refers to this thing or means this operation” without letting the desired interpretation manufacture its own confirmation.

Quick Read

Direct answer: Voynich semantics is not recovered by attaching plausible words to recurring forms. A strong semantic claim should climb an evidence ladder: observable structure → stable functional role → independently constrained referent → lexical or operational meaning → connected interpretation. Each transition requires new evidence. Images, dictionaries, historical comparators and statistical embeddings are useful only when they validate a value independently rather than supplying the value and then being counted as confirmation.

A real statistical class does not automatically have a known semantic class.
A token that behaves like a noun is not yet a known noun.
A label attached to an image proves spatial relation, not necessarily naming.
A repeated visual motif can encode identity, class, operation, state or decoration.
A distributional embedding measures contextual similarity, not lexical meaning directly.
Images can ground meaning only when the textual prediction was fixed without using the same image.
Dictionary matching is especially dangerous when segmentation, sound values and language are flexible.
Multiple-testing and “dictionary collision” can make accidental lexical matches look impressive.
Semantic values should propagate: the same proposed value should constrain other occurrences.
A successful value should reduce future ambiguity rather than require new exceptions.
Cross-section recurrence can be strong grounding evidence if the same textual identifier tracks independently matching visual entities.
Cross-reference claims therefore become especially important downstream.
Cipher models can preserve semantics while disguising surface forms.
Notation models can carry functions or classes without ordinary lexical words.
No accepted Voynich semantic lexicon currently satisfies the full manuscript-wide grounding burden.

The Five-Layer Evidence Ladder

A disciplined semantic analysis can be organised into five layers.

Layer 1: Observation

A pattern exists.

one token occurs frequently;
one ending prefers line ends;
one label family occurs near stars;
one visual root recurs on two pages.

This is the strongest floor because another observer can check it.

Layer 2: Functional role

The pattern performs a stable structural job.

A token class consistently separates entries.

An ending selects a predictable neighbouring class.

A label family acts as identifiers rather than prose.

We still do not know the worldly meaning.

Layer 3: Reference

The form can now be tied to something external or visually independent.

The same identifier follows the same object across two pages.

A ring position aligns reproducibly with one external historical catalogue.

A label value predicts an image class it was not trained on.

Layer 4: Semantic value

Now we can begin to say what the referent or operation means.

“rose”.

“heat”.

“second degree”.

“and”.

Layer 5: Connected interpretation

Values interact under one grammar or procedure to recover larger content.

This is where translation or operational reading begins.

Most Voynich claims jump from Layer 1 or 2 directly to Layer 4.

That jump is the semantic gap.

Structure Is Necessary but Not Sufficient

Suppose a clustering algorithm divides Voynich tokens into four robust classes.

Excellent.

Those classes might be:

nouns, verbs, adjectives and particles;
ingredient, operation, quantity and state codes;
four cipher-table families;
four positional templates;
four generator states.

The same clustering can be real under several semantic ontologies.

Clustering discovers separability.

Grounding discovers what the separation is about.

Distributional Semantics Is Not a Dictionary

Modern computational linguistics often represents words by the contexts in which they occur.

Words used similarly become close in a vector space.

This can reveal grammatical or topical classes.

But a vector does not tell us that one cluster means “plants”.

It tells us that its members behave similarly.

Recent public computational projects report productive morphological paradigms, preliminary lexica and cross-section label networks.

Those are interesting hypotheses.

The grounding question remains:

what independent evidence ties this distributional neighbourhood to that historical referent?

Why Pictures Are Both Our Best Hope and Our Biggest Trap

Voynich is multimodal.

This is unusual and valuable.

A label sits beside a plant.

A word occurs beside a star.

A paragraph wraps around a vessel.

Images give us external structure the text alone does not.

But they also make circular confirmation effortless.

See a plant.

Guess “root”.

Find a nearby token.

Declare that token “root”.

Then say the translation is confirmed because the page contains a root.

The image supplied both the hypothesis and the confirmation.

No independent evidence was added.

Cross-Validation Requires Direction

To use pictures properly, information must flow one way during hypothesis construction and the other way during validation.

Example:

Infer a textual class using only running text.
Freeze the class and its predicted function.
Apply it to labels on unseen pages.
Only then inspect the images.
Ask whether the predicted class aligns with an independently coherent visual category.

Now text predicted image.

The agreement adds information.

Visual Semantic Profiling Is a Useful New Control

A 2026 preprint by Jacob Lyons explores visual semantic profiling of Voynich page images using machine-learning representations and section discrimination.

The public companion repository releases profile vectors, statistical artefacts and reproduction scripts.

This type of work is valuable because it can quantify visual separation without first naming every image.

But a visual classifier distinguishing herbal from zodiac does not automatically recover what the text says on either page.

Visual structure becomes semantic evidence only when it constrains textual predictions independently.

The Label Problem Is Grounding in Miniature

Labels are short and anchored.

This makes them tempting.

A label beside a star could be:

star name;
number;
degree code;
class;
person name;
mnemonic handle.

Placement reduces the hypothesis space.

It does not reduce it to one.

The Zodiac Sigla Problem is powerful precisely because it asks about label capacity and list structure before claiming the external names themselves have been found.

Capacity Comes Before Meaning

Suppose a label can carry only enough information for two underlying letters under a proposed cipher model.

Then a theory claiming that label literally spells a nine-letter star name is incompatible with the channel.

This is a strong pre-semantic constraint.

Capacity analysis can eliminate meanings before choosing them.

That is exactly what good grounding needs.

The Dictionary-Collision Problem

A large historical dictionary contains many words.

Allow:

uncertain vowels;
several sound values per glyph;
optional letters;
abbreviations;
spacing changes;
several target languages.

Now almost every short Voynich string can be made to resemble something.

The correct question is not:

Can I find a matching word?

It is:

How often would a match this good appear under the same search freedom if the token were unrelated?

That is why dictionary-collision controls matter in the newest shorthand and language-identification work.

Semantic Values Should Reduce Entropy of Future Interpretation

A good discovered value is expensive at first and cheap later.

Suppose one token truly means “water”.

That value now constrains every future occurrence.

It should help interpret neighbours.

It should make some readings impossible.

The dictionary should converge.

If the same token later needs to mean “sun”, then “woman”, then “boil”, the first grounding did not constrain the system.

A real semantic discovery should make the next page harder to interpret incorrectly.

Function Before Lexical Meaning

Sometimes we can know what a sign does before what it means.

A star-shaped mark begins entries.

A gallows-rich form marks paragraph starts.

A suffix-like class predicts a following class.

Those are functional statements.

They can be scientifically strong without lexical translation.

This is often the safer bridge toward semantics.

Notation Makes Functional Grounding Especially Important

If Voynich is technical notation, some units may never correspond to ordinary words.

They may mean:

repeat;
class A;
state B;
apply operation C;
look elsewhere.

The Semantic-Grounding Problem therefore cannot assume every successful value belongs in a dictionary.

Some may belong in an instruction manual.

Cipher Makes Surface Grounding Dangerous

If visible forms are homophonic ciphertext, two different tokens can mean the same source word.

The same visible component may also be only part of an encoding group.

Grounding the ciphertext directly to images can therefore mistake encoding mechanics for semantics.

A cipher model should first recover the equivalence classes.

Then semantics is attached to the underlying source unit.

Currier A/B Is a Grounding Stress Test

Suppose one A-form is grounded as “root”.

What happens on B pages?

If the same concept is present but surface forms change, one transformation should map them.

If A and B require completely separate dictionaries, perhaps they are different languages or systems.

Either way, semantic grounding should clarify Currier rather than ignore it.

Genre Is a Semantic Prior, Not a Semantic Verdict

If Voynich is a medical reference book, “root”, “water”, “heat” and “dose” become plausible vocabulary.

But genre only changes prior probability.

It does not allow us to assign those words to arbitrary tokens.

The newly published Genre Problem should narrow the candidate world, not fill the dictionary.

Audience Is Another Semantic Prior

A specialist audience can tolerate short identifiers and compressed notation.

A novice audience requires explicit labels and explanatory grammar.

This helps decide whether a short Voynich label is likely to be a full name, abbreviation or code.

Again, audience constrains meaning without determining it.

Cross-References Can Become Powerful Grounding Anchors

Suppose the same rare label occurs beside two visually matching plant roots on distant pages.

This is stronger than one local label-image match.

The textual recurrence and visual recurrence are independent dimensions.

If the relationship repeats across many pairs, the label begins to act like an identifier.

We still may not know the plant’s real-world species.

But we may know that the manuscript treats two images as the same internal entity or class.

This is why the Cross-Reference Problem is the next article in this batch.

Identity Grounding Can Precede World Grounding

This distinction is subtle and useful.

We may discover:

token X refers to the same thing on pages 10 and 90.

without knowing what that thing is in modern taxonomy.

Internal identity is already semantic progress.

World identity comes later.

A Semantic Claim Must Have a Failure Condition

If token X means “water”, what would falsify it?

X appearing routinely as isolated plant labels?
X occurring in contexts incompatible with liquid classes?
another independently grounded token already filling the same role?
Currier transformations breaking the value?
a historical cipher inversion mapping X elsewhere?

If nothing can falsify “water”, then “water” is decoration added by the researcher.

Negative Results Are Semantic Progress

Retained Voynich work has already rejected several attractive semantic totalisations.

universal container-operation code;
zodiac as established master timing control;
strict plant→preparation→body→zodiac→recipe pipeline;
single totalising programme map.

These failures are important because they remove semantic freedom.

Every failed legend narrows what future meanings are allowed to do.

What Survives the Semantic-Grounding Work

Voynich contains abundant reproducible structure.
Structure alone does not identify meaning.
Functional roles can often be inferred before lexical values.
Images provide powerful independent evidence only when they are not used to invent the textual value being validated.
Distributional similarity supports equivalence classes, not dictionary entries by itself.
Dictionary search requires explicit collision and multiple-testing controls.
Cross-reference networks can ground internal identity before real-world identity.
A valid semantic value should propagate and reduce future interpretive freedom.
Cipher, shorthand and notation alter the layer at which semantics should be attached.
No accepted Voynich semantic dictionary currently survives the full burden.

What Does Not Survive as Established Knowledge

A robust cluster has a known meaning.
A label literally names the adjacent picture.
A visually plausible plant identification grounds its label automatically.
A high cosine similarity reveals a lexical synonym by itself.
A dictionary hit proves translation.
A repeated vessel shape has a decoded operation value.
A successful genre hypothesis supplies the vocabulary.
A functional notation value can be invented from page appearance and then validated by that same appearance.

A Better Semantic-Grounding Analysis

Start with reproducible structure.
Infer functional role without assigning worldly meaning.
Freeze the structural model.
Use an independent channel—image, historical catalogue, cross-reference or external crib—to constrain reference.
Quantify search freedom and dictionary-collision probability.
Assign a semantic value only after reference is constrained.
Propagate the value to every occurrence under fixed rules.
Test Currier, hand and document-role transfer.
Require the value to predict unseen contexts.
Preserve every contradiction rather than renaming it an exception automatically.

What Would Count as a Real Semantic Breakthrough?

Imagine a rare Voynich label family is identified distributionally without using images.

The family predicts a specific document role.

The rule is frozen.

On unseen pages, the same family occurs beside a visually coherent object class.

Several distant occurrences align with the same internal image identity.

An independent historical comparator then supplies one plausible real-world referent.

A separate textual mechanism fixes the sound or code value without using that comparator.

Both routes converge on the same meaning.

The value then predicts neighbouring grammar on new folios.

That would be semantic grounding.

Meaning becomes evidence when independent routes converge on the same value and that value makes future interpretation less free.

Primary School: Evidence and Meaning Are Different Boxes

Show a child an unfamiliar symbol beside a picture of a dog.

Ask what the symbol means.

“Dog” is a good guess.

Then reveal that the same symbol appears beside several cats and rabbits too.

Perhaps it means “animal”.

One more occurrence beside a veterinary medicine bottle changes the possibilities again.

The lesson is simple: context narrows meaning; one context rarely finishes it.

Secondary School: Predict Before Looking

Give students an invented language and several hidden pictures.

Let them infer word classes from text first.

Freeze their predictions.

Then reveal the pictures.

Students experience the difference between independent validation and guessing from illustrations.

JC and Adult Readers: Grounding as Identifiability

At a higher level, semantic grounding is an identifiability problem.

Several latent semantic systems can generate the same structural observations.

Independent modalities reduce the equivalence class only when their errors and hypothesis-generation processes are genuinely independent.

The goal is not maximum interpretive richness.

It is the smallest semantic model forced by the greatest number of independent constraints.

Reader Checklist: Before You Believe a Voynich Meaning

What exact observation is being interpreted?
Has a stable functional role been shown first?
Was the image used to invent the meaning?
If yes, is the image also being counted as confirmation?
How many dictionary candidates were searched?
How flexible are segmentation and sound values?
Does the value apply to every occurrence?
Does it survive Currier and document-role changes?
Does it reduce ambiguity on unseen text?
What observation would falsify it?
Is the claimed value lexical, functional or referential?
Are those levels being kept separate?

Frequently Asked Questions

Do Voynich statistical clusters have meanings?

They may, but clustering establishes structural similarity before semantic identity. Additional independent evidence is needed to ground a class.

Can pictures tell us what nearby words mean?

They can constrain possibilities strongly, especially for labels, but the safest evidence comes when the textual value was predicted independently before the picture was consulted.

Why are dictionary matches weak?

Because large dictionaries plus flexible sound mappings, segmentation and candidate languages create many accidental matches. Search freedom must be quantified.

Can we know internal identity without knowing the real-world object?

Yes. Repeated labels and matching visual entities can potentially show that the manuscript treats two instances as the same internal entity or class even before modern identification is known.

What is the strongest current conclusion?

Voynich structure is increasingly measurable, but the bridge from structure to stable semantic value remains one of the central unsolved problems.

Research and Further Reading

The Final Idea

The hardest part of Voynich may not be finding patterns anymore.

Computers can find patterns everywhere.

The hard part is making one pattern answer to the world.

This sign does not merely recur.

It refers.

This ending does not merely correlate.

It performs the same grammatical operation.

This label does not merely sit beside a picture.

It tracks the same entity somewhere else.

Once those bridges begin surviving independently, meaning will stop being something we add to Voynich.

It will become something the manuscript itself refuses to let us replace with another meaning.

Voynich | Everything eduKate Knows and Tested | The Semantic-Grounding Problem: How Does Structure Become Meaning?