The Voynich Manuscript has been divided into “sections” for generations.
Herbal.
Astronomical or astrological.
Biological or balneological.
Cosmological.
Pharmaceutical.
Recipes or starred text.
Those names are convenient.
They are also interpretive.
Now imagine removing the transcription entirely.
No EVA.
No word counts.
No Currier frequencies.
Only the page images remain.
Can a computational system recover those scholarly section labels from visual information alone?
Jacob Lyons’s 2026 preprint Visual Semantic Profiling of the Voynich Manuscript reports that it can: a multinomial logistic regression over sixteen visual-semantic profile dimensions recovered the scholarly section labels with 90.4% leave-one-out accuracy on the paper’s primary 197-page dataset.
The result is striking.
It is also easy to overstate.
A machine that can recognise the Voynich sections has shown that the illustrations are structurally differentiated. It has not shown that the labels we give those sections are their historical meanings.
Quick Read
- Lyons’s 2026 work is a preprint with a public companion repository and Zenodo deposits.
- The study treats page images as the primary signal and does not attempt to decipher Voynichese.
- It profiles analysable pages across sixteen human-authored visual-semantic dimensions using a frozen vision-language model.
- All sixteen dimensions reportedly differ significantly across scholarly sections under several statistical tests.
- A leakage-controlled multinomial logistic regression reports 90.4% leave-one-out section-classification accuracy on 197 unambiguous pages.
- The paper reports 92.4% accuracy using the raw foundation-model embeddings and 72.1% using six handcrafted layout features.
- The layout result shows that page geometry alone already contains substantial section information.
- The raw embeddings outperform the interpretable sixteen-dimensional projection slightly, showing that some discriminative visual information is lost when compressed into the human-authored lens.
- The scholarly section taxonomy is itself based heavily on visible imagery, so high visual recoverability is not equivalent to recovering hidden semantics.
- The strongest conclusion is that the manuscript’s illustration programme contains robust machine-measurable page-family structure.
- The study does not establish that the pictures mean the same thing as the surrounding text.
- It does not identify authorship, language, provenance or decoded subject matter.
Why This Experiment Is Worth Doing
Most computational Voynich work begins by removing the pictures.
Researchers want a clean text corpus.
Illustrations become noise masks, layout regions or obstacles to OCR-like segmentation.
That is understandable.
It also ignores half the manuscript’s information architecture.
Voynich is not a plain-text document.
Plants occupy pages.
Human figures sit inside pools.
Stars become labels.
Containers organise fragment rows.
Circles determine where text can run.
A visual-only experiment therefore asks a legitimate independent question:
how much manuscript organisation remains recoverable after the unreadable script is deliberately ignored?
What the Model Actually Sees
The study uses a frozen vision-language foundation model.
It maps page images and natural-language descriptors into a shared embedding space.
Sixteen human-authored archetype dimensions are then used to create an interpretable page profile.
This matters because the representation is not neutral.
The model contributes knowledge inherited from its training corpus.
The researchers contribute the descriptor vocabulary.
The page contributes its visible form.
The resulting vector is therefore an interaction among all three.
The paper explicitly acknowledges this.
That acknowledgement is essential because “zero-shot” does not mean “interpretation-free”.
Why 90.4% Is Impressive
The reported classifier is not merely re-scoring the same page it trained on.
The main result uses leave-one-out cross-validation with the preprocessing fitted inside the pipeline, reducing a common leakage route.
The paper reports:
- 90.4% leave-one-out accuracy;
- a Wilson 95% confidence interval around 85.4–93.7%;
- a permutation-test probability below 0.001;
- a 20% chance baseline;
- a majority-class baseline around 59.9%.
Under the paper’s labels and representation, the page families are therefore strongly separable.
This is not a tiny effect hiding behind p-values.
The classifier is recovering something large and persistent in the visual channel.
Why 90.4% Is Not a Decipherment
The scholarly section labels are themselves based heavily on visible imagery.
A page with one enormous plant is called herbal because it has an enormous plant.
A page full of nude figures in pools is called biological or balneological because of those visible figures and pools.
A page full of stars is classified partly because of the stars.
So a visual model recovering these labels is not like a blind machine discovering a hidden taxonomy from nothing.
It is partly recovering categories humans already defined from visual evidence.
This does not make the result trivial.
It changes what the result means.
The experiment validates the robustness of the visual partition more strongly than it validates the semantic names attached to that partition.
The Layout Ablation Is Especially Important
The study reports about 72.1% classification accuracy using only six handcrafted layout features.
This is one of the most useful controls in the paper.
It tells us that a large fraction of section separability exists even without rich semantic embeddings.
Page architecture alone matters.
- amount of drawing;
- arrangement of text and image;
- circularity;
- figure density;
- object distribution;
- spatial composition.
This connects directly with When Text Meets Image.
Voynich sections are not merely thematic clusters.
They are page-design regimes.
The Raw Embedding Result Is Another Useful Control
The paper reports about 92.4% accuracy using raw 768-dimensional foundation-model embeddings.
That slightly exceeds the sixteen-dimensional interpretable profile.
This tells us two things.
- The visual signal is not created entirely by the chosen sixteen descriptors.
- The interpretable projection sacrifices some discriminative power in exchange for a representation humans can discuss.
This is a healthy trade-off.
Interpretability should be treated as a design choice rather than a guarantee of truth.
All Sixteen Dimensions Discriminate—But That Does Not Mean All Sixteen Are Historical Categories
The paper reports extremely strong section differences across all sixteen profile dimensions under ANOVA, Welch and Kruskal–Wallis tests.
That tells us the page families occupy very different visual-semantic regions under the model.
It does not prove medieval makers conceptualised those exact sixteen dimensions.
The dimensions are modern analytical lenses.
A medieval illustrator did not need to think in those variables for the page to score differently on them.
This distinction matters especially in AI-based cultural analysis.
Machine-readable differences can be real without the machine’s conceptual vocabulary being historically native.
Can Pictures Carry Meaning When Text Is Unread?
Yes, in a limited sense.
A plant drawing differs from a bathing scene.
A circle of zodiac figures differs from a row of containers.
A machine can measure those differences.
But the word meaning contains several levels.
- visual category: plant-like page versus circular diagram page;
- document function: reference list versus procedure versus illustration;
- historical subject: medicine, astrology, botany, cosmology;
- specific semantics: this plant is X, this figure is Y, this sentence says Z.
The classifier strongly supports the first level.
It may help investigate the second.
It does not establish the third or fourth by itself.
Why the Six Sections May Still Not Be Six Original Sections
The existing Six Sections That May Not Be Six Sections article remains important.
High classification accuracy does not prove the modern six-section taxonomy equals the manuscript’s original conceptual organisation.
Why?
- Visual regimes can overlap conceptually.
- A plant page and a pharmaceutical page may belong to one larger technical system.
- A zodiac page and a human-body page may participate in one astro-medical programme.
- Current page order is disturbed.
- Missing leaves can split or hide transitions.
The classifier can recover the visible clusters we hand it while the historical ontology remains different.
The Stronger Future Test: Predict Text Regime From Images
A more independent experiment would cross channels.
Train a visual model without text.
Then ask whether its visual clusters predict:
- Currier A/B;
- hand assignment;
- token-family distribution;
- hapax rate;
- edge coupling;
- line-position behaviour;
- learned unit inventory.
If image-derived states predict independent text properties on held-out pages, the visual channel begins to tell us something deeper than “plants look like plants”.
That is where multimodal Voynich research becomes genuinely powerful.
The Reverse Test Is Just as Important
Can the text predict the picture family?
Montemurro and Zanette already showed that local vocabulary carries section information.
A modern held-out classifier can push this further.
Train on textual features.
Predict visual section.
Then compare text→image and image→text errors.
Pages where both channels disagree with the conventional label may be especially valuable.
They may reveal transitional pages, misclassification, mixed document roles or disturbed order.
Out-of-Distribution Controls Matter
The public repository includes a small out-of-distribution probe using Tacuinum Sanitatis pages.
This is a useful instinct.
A visual semantic system should not merely find “medieval manuscript-ness” and mistake it for Voynich section structure.
Future work should greatly expand this control set.
- known herbals;
- astronomical manuscripts;
- balneological books;
- medical recipe collections;
- mixed encyclopaedias;
- unrelated decorated codices.
The question is whether Voynich occupies distinctive regions or simply reproduces broad medieval visual genres.
The Foundation Model Is Not a Neutral Eye
Vision-language models inherit their training worlds.
They know modern captions.
Modern categories.
Modern image distributions.
They may recognise “plant”, “diagram”, “human figure” and “container” partly because those categories are abundant in modern labelled data.
This can still be useful.
It means the result should be framed as:
modern machine-readable visual structure in the manuscript.
Not:
the medieval makers’ own semantic categories recovered directly.
The Proprietary-Pipeline Limitation
The paper and repository disclose substantial outputs, profile vectors, statistics and reproduction scripts.
They also state that some implementation details of the profiling method are covered by a pending provisional patent and are not fully public.
This creates a reproducibility distinction.
- The released profile vectors and downstream analyses can be reproduced publicly.
- Exact regeneration of those vectors from raw images depends on implementation details not fully disclosed in the public paper.
That does not invalidate the reported classifier.
It limits how independently the full pipeline can currently be reconstructed from first principles.
Preprint Status Matters
The repository identifies Zenodo as the preprint of record.
An arXiv submission was rejected at moderation under a stated category-policy issue, and the work has not yet become peer-reviewed consensus.
That status should be reported plainly.
Moderation rejection is not scientific falsification.
Zenodo publication is not peer review.
The correct status is:
reproducibility-oriented preprint with substantial released downstream data, awaiting wider independent evaluation.
What the Visual Classification Problem Does Not Prove
- It does not decipher Voynichese.
- It does not prove the scholarly section names are historical meanings.
- It does not prove the six modern sections are six original chapters.
- It does not prove image and text carry identical semantics.
- It does not identify provenance.
- It does not identify authorship.
- It does not eliminate layout as a major source of classifier success.
- It does not make a foundation model a neutral medieval observer.
What It Can Give Us
- Strong evidence that Voynich page families are visually machine-separable.
- A quantitative measure of how much section structure lives in layout versus richer visual features.
- An interpretable visual feature space for comparing pages.
- A route to cross-channel prediction between imagery and text statistics.
- A way to identify mixed or anomalous pages that do not fit their conventional section.
- A computational framework for testing whether visual structure generalises to known medieval comparators.
Reader Checklist
- What labels is the classifier asked to recover?
- How were those labels originally defined?
- Are they already visually obvious by construction?
- What is the chance and majority baseline?
- Was preprocessing kept inside cross-validation?
- How well do layout-only features perform?
- How well do raw embeddings perform?
- Are the semantic descriptors human-authored?
- Can the full image-to-profile pipeline be independently reproduced?
- Do results generalise to external medieval manuscripts?
- Does the visual representation predict independent text properties?
- Is classification being described as semantic decipherment?
Research Foundations
- Jacob Lyons — Visual Semantic Profiling of the Voynich Manuscript, public companion repository.
- Public preprint source.
- Zenodo preprint of record.
- Released per-page profile dataset.
- The Six Sections That May Not Be Six Sections.
The Final Idea
The pictures do not have to translate the text to matter.
They can carry their own structure.
The 2026 visual-classification result makes that structure measurable at scale.
The next step is harder.
Make the visual structure predict something independent.
A classifier that recognises the sections tells us the manuscript has visual organisation. A model that predicts unseen textual behaviour from those images would begin to tell us how deeply that organisation runs.