eduKateSG · VOYNICH RESEARCH LIBRARY · MATHEMATICS & SCIENCE SYNTHESIS
Voynich | The Mathematics and Science
What an unread manuscript taught us about measurement, uncertainty, information, matter, structure—and the discipline of not knowing.
← Enter the Voynich Research Library
There are two careless ways to approach the Voynich Manuscript.
The first arrives carrying a key. It sees a symbol and calls it a letter. It sees a space and calls it a word boundary. It sees a round diagram and calls it astronomy. It sees a root-like form and calls it botany. It finds a city in which mathematics, medicine and natural philosophy once flourished, and calls that city the origin.
The second careless approach arrives with a shrug. It says the manuscript is unreadable, therefore nothing can be learned. It mistakes the absence of translation for the absence of information.
Both approaches are wrong.
The Voynich Manuscript is usually presented as a locked door. That metaphor is irresistible, but it quietly damages the inquiry. A locked door implies that the correct key already exists and that the investigator’s task is simply to find it. Every new decipherer therefore arrives holding an object shaped like an answer.
A better metaphor is a damaged warehouse. The labels have disappeared. Some aisles may have been rearranged. Several rooms may have been constructed for different purposes. Parts of the inventory may be missing. Later custodians left numbers and traces of handling. The surviving packages carry repeated marks, but we do not yet know whether those marks identify contents, instructions, quantities, sounds, operations, categories—or something for which our modern vocabulary is poorly fitted.
Yet the warehouse is not empty. Its walls can be surveyed. Its materials can be dated. Its circulation paths can be mapped. Repeated packages can be counted. Workstations can be distinguished. Structural transitions can be measured. Proposed labels can be withheld until the geometry survives a blind test.
The manuscript has not yet yielded its meaning. It has yielded a curriculum in how meaning must be earned.
The most interesting mathematics and science did not arrive after decipherment. They arrived because decipherment remained unavailable. We learned how to measure surprise without calling it language; how to detect association without calling it translation; how to date parchment without pretending we had dated the act of writing; how to distinguish hands without naming their owners; how to recover recurring visual structures without deciding what they depict; and how to place a manuscript inside a historical field of possible capabilities without converting possibility into provenance.
Perhaps most importantly, we learned that a responsible investigation is not defined by how quickly it produces a theory. It is defined by how difficult it makes that theory to survive.
The Object Before the Mystery
The Voynich Manuscript is Beinecke MS 408, an illustrated parchment codex written in an unidentified script by an unknown author. Its surviving physical construction includes numerous double and multiple foldouts, difficult collation, missing leaves and an uncertain original arrangement. Those facts matter because the manuscript is not merely a string of symbols. It is a constructed object whose folds, bifolios, quires, losses and later handling constrain every account of its sequence.
Radiocarbon testing of four parchment samples placed the material in the early fifteenth century, conventionally reported as a 95 per cent interval of approximately 1404–1438. That does not give the manuscript a birthday. It dates the biological material from which the sampled parchment was made; it does not by itself identify the writing date, the place of inscription, the author, the language or the manuscript’s purpose.
A page has more than one clock: animal life, parchment preparation, inscription, colouring, assembly, binding, foliation, damage, repair and ownership. These clocks may be close together. They are not automatically identical.
Scientific rule: before every date, ask what exactly has been dated?
External references: Yale Beinecke Library · Voynich radiocarbon documentation
The First Warehouse: Measurement Before Meaning
A glyph is an observation; a letter is an interpretation
The easiest mistake in Voynich research is also the most foundational: we see a recurring mark and call it a letter. A glyph is an observed shape or a class of shapes. A letter is a claim about how that shape functions inside a linguistic writing system. A glyph might represent a letter, syllable, abbreviation, number, morpheme, control sign, cipher group, ornamental variation, scribal gesture or component of a larger unit.
The same danger applies to spaces. A visible gap may be a lexical boundary, but it might also mark abbreviation, phrasing, graphic convenience, line planning, cipher grouping, prosodic segmentation or a weak junction inside a larger unit.
A recent August 2026 preprint tested three assumptions directly: glyph equals letter, blank-delimited token equals word, and every visible separator equals a word space. The authors reported that none of the three identifications survived their comparison framework and argued instead for recurrent multi-symbol units, weak adjacent token-identity order, stronger information at token edges and more than one separator regime. Because this is a recent preprint, its estimates deserve independent replication. Its methodological challenge, however, is powerful: unit names must follow validation, not precede it.
Observed unit ≠ inferred function.
A mark is not yet a phoneme. A token is not yet a word. A picture is not yet an object identification. A cluster is not yet a topic. A statistically significant boundary is not yet a grammatical rule.
See: A Glyph Is Not a Letter, a Token Is Not a Word, a Space Is Not a Space.
Transcription is an instrument, not a transparent window
To count the Voynich script, researchers usually work from a transcription: a modern symbolic representation of what someone saw on the page. That transcription is indispensable, but it is not the manuscript itself. It is a measurement instrument placed between the manuscript and the analysis.
Recorded data = manuscript signal + reader decisions + imaging limits + coding conventions.
A transcriber decides whether two adjacent strokes form one glyph or two, whether a mark is a variant or a separate class, whether a narrow gap is certain or uncertain, and how to represent damage or unusual forms. Every one of those choices may alter later counts.
A frequency table may therefore be mathematically exact while remaining conditional on a debatable segmentation. The serious question is not merely what result we obtain from one transcription, but whether that result survives another transcription, another segmentation policy, another glyph inventory and another reader.
The Mathematics Warehouse
Counting was useful—but counting was never enough
The first computational approaches to an unknown text naturally count things: symbols, apparent tokens, distinct types, type frequencies, token lengths, positional preferences and regional distributions. These are useful questions because they do not require translation.
But every count has hidden dependencies: Count = f(unit definition, sample boundary, classification rule). Change the unit and the count changes. Change the transcription and the count changes. Split one frequently used token into two components and the vocabulary changes.
A number without its counting rule is therefore not evidence. It is decoration. The denominator, segmentation policy and uncertainty class must travel with every number.
Frequency laws describe distributions, not causes
Voynichese displays non-uniform frequency: some forms recur often; many appear rarely. Researchers have compared those distributions with patterns seen in known languages. That is interesting. It is not a translation.
Natural language can produce long-tailed distributions. Formulaic notation can. Ciphers can preserve or transform them. Copy-and-modify processes and deliberately constrained pseudo-text can produce them too. The mathematical task is not to reward one familiar-looking statistic. It is to find a joint profile that competing explanations cannot easily reproduce.
Entropy: the mathematics of surprise
Shannon entropy measures average uncertainty or surprise in a distribution:
H(X) = − Σ p(x) log₂ p(x)
Conditional entropy asks how much uncertainty remains about the next symbol when something about the current state is known:
H(Xₜ₊₁ | Xₜ)
Voynichese is famous for being unusually predictable at the character level. Luke Lindemann and Claire Bowern compared Voynich transcriptions with hundreds of modern-language samples and historical texts. They reported that major changes to glyph composition were not sufficient to align Voynich conditional entropy with their comparison languages; character placement remained unusually constrained within apparent words.
This is a real mathematical observation. It does not mean the manuscript is simple, meaningless or solved. Low conditional entropy can arise from restricted scripts, positional rules, abbreviations, repeated formulae, multi-symbol encoding, constrained production, artificial notation or a cipher operating at a different scale from the one being measured.
The deeper lesson is that entropy is representation-sensitive. A musical score measured as isolated ink strokes has one entropy profile; measured as notes, chords, phrases or harmonic functions, it has others. A chemical formula measured as individual characters differs from the same formula measured as element symbols or reactions.
The Voynich question is therefore not only “how predictable is the sequence?” It is “at what level does the sequence become properly represented?”
See: Lindemann & Bowern, Character Entropy in Modern and Historical Texts · Why Voynichese Is So Predictable · Frequency Laws.
Information is not meaning
Mutual information measures how much knowing one variable reduces uncertainty about another. If the final glyph of one apparent token helps predict the first glyph of the next, that is evidence of dependence. It does not tell us why the dependence exists.
The dependency might be linguistic agreement, a scribal joining preference, cipher state, line construction, a copy-and-modify process or a modern segmentation that has cut one larger unit into two.
Mutual information ≠ shared meaning.
Information theory is valuable precisely because it can detect organisation before semantics. The discipline is to stop where the measurement stops.
Long-range structure: the text remembers more than its neighbour
Statistical studies using information theory, network analysis and intermittency have reported long-range and non-random structure in the manuscript. A form may appear in bursts rather than uniformly; some forms co-occur; distributions can change across manuscript regions. An unread text can therefore possess geography before we know the names of its districts.
But organisation is not unique to natural prose. Catalogues, tables, ciphers, ritual sequences and pseudo-text generators can all be organised. A long-range signal strengthens the case that there is an underlying process. It does not identify that process.
Topic models: clustering is not reading
A 2021 study applied latent Dirichlet allocation, latent semantic analysis and non-negative matrix factorisation to Voynich pages. Its authors reported that computational clusters corresponded closely to a conjunction of proposed scribal hand and illustration-defined subject groupings. That is intriguing, but the name topic model can mislead. The mathematics identifies latent distributional components. It does not open the manuscript and discover a cupboard labelled “botany.”
A cluster can be driven by subject matter, scribal habit, local notation, page layout, copying source, production stage, cipher settings, formulaic templates—or several at once.
See: Sterneck, Polish & Bowern, Topic Modeling in the Voynich Manuscript.
The newest warning: familiar structure is not discriminating structure
The 2025 Naibbe cipher is a useful modern comparator. Michael Greshko designed a hand-workable verbose homophonic substitution system capable of turning Latin and Italian into Voynich-like ciphertext. The paper does not claim that Naibbe is the manuscript’s actual cipher. Its scientific value is experimental: it demonstrates that a historically plausible construction can reproduce several striking Voynich-like properties while still failing to reproduce others.
This changes the evidential weight of old observations. Low entropy remains real. Its power to discriminate among hypotheses becomes smaller when more than one mechanism can produce it.
A feature is scientifically useful not merely when it is striking, but when plausible alternatives struggle to generate it under fair conditions.
The State Warehouse: Position Is Information
Voynich forms are not distributed indifferently across the page. Some glyphs or glyph families prefer initial, medial or final positions. Some forms behave differently near line boundaries. Paragraph-initial lines can differ from ordinary lines. These effects suggest that the manuscript’s production system carries state.
A simple statement is: P(Xₜ₊₁) ≠ P(Xₜ₊₁ | Sₜ). Here, the state may include line position, previous terminal form, manuscript region, scribal hand, page interface, illustration context, boundary strength or an unobserved production condition.
The important point is that state dependence can be detected before the state is understood. Once a researcher sees a boundary effect, it is tempting to name it—“begin,” “apply,” “select,” “instruction mode.” Those names are hypotheses, not measurements.
An odds ratio is not a dictionary
If an apparent form occurs more frequently after one boundary class than another, an odds ratio can summarise the association. An odds ratio above one indicates a difference in likelihood under the two conditions. It does not translate the event.
Within the current eduKate text-state lane, a boundary-associated effect was reproduced across two transcription regimes with strong effect sizes. The authorised statement remains narrow: a local form-selection difference appears associated with boundary state. It is not licensed as a lexical meaning.
A large effect size can coexist with a small authorised claim.
Line beginnings may behave like resets
In many systems, boundaries reset state: computer code, musical phrases, manufacturing cycles, poetic lines and some cipher procedures. Voynich line beginnings show distinctive distributional behaviour. This makes “reset” a useful model, but still only a model. The scientific task is to test whether a reset model predicts held-out observations better than line planning, scribal ornament, copy-text geometry or a generator with positional preferences.
The manuscript turns a visible line into a statistical question: line as appearance is observable; line as functional unit must be earned.
The Matrix Warehouse: The Manuscript Does Not Have One Structure
The conventional visual names—herbal, astronomical, balneological, cosmological, pharmaceutical and recipes—are useful catalogue aliases. They become dangerous when a visual alias is promoted into a proven semantic department.
The pages can be divided by many axes: physical quire, bifolio, proposed scribal hand, Currier text regime, visual morphology, prose-versus-label geometry, distributional cluster, foldout status, line architecture and illustration density. These divisions do not always change together.
A scribal change can occur inside one visual family. A distributional change can occur without a physical break. A quire can contain more than one page morphology. Two pages can look visually similar while behaving differently in their text distributions.
What the frozen Test 1b matrix taught us
The frozen eduKate Test 1b matrix recorded 227 page units and 52 represented bifolios, with multiple physical, scribal, visual and distributional classifications. The strongest surviving result was not a single semantic map or unified workflow. It was an overlapping multi-axis object. Proposed hand coherence, Currier-regime coherence and neutral visual-morphology coherence were much stronger than distributional-cluster coherence across complete comparable bifolios.
Those figures do not decode the manuscript. They teach something more fundamental: physical production, visual design and text distribution are related, but they are not interchangeable.
Association does not collapse categories
A contingency table can test whether two classifications are associated. Cramér’s V can summarise the strength of that association. But a strong association between visual class and text distribution would not make them the same variable.
Weather and umbrella use are associated. Umbrellas are not weather. A page morphology and a text regime can align. The morphology is not thereby the meaning of the text.
The right model may be a hypergraph
A page can belong simultaneously to a physical bifolio, a quire, a scribal-hand class, a visual family, a text regime, a distributional cluster and a particular interface type. No single membership owns the page. The page exists at their intersection.
The apparent mess is not necessarily a failure of classification. It may be the correct shape of the object.
Related: Hand × Bifolio × Section: Reconstructing the Production Matrix.
The Geometry Warehouse: Components Before Objects
The manuscript’s illustrations are seductive because the human visual system is an identification machine. We see branching and think plant. We see tubes and think plumbing or anatomy. We see circles and stars and think astronomy. We see walls and towers and think map or city.
A safer analysis begins with geometry: axes, branch events, terminal forms, radial symmetry, repeated lobes, bands or collars, nested components, connectivity, negative space and attachment geometry.
Topology survives deformation
Geometry measures length, angle, curvature, area and proportion. Topology asks what remains unchanged when a form is stretched or deformed without cutting or joining. A loop remains a loop. A branch point remains a branch point. A connected component remains connected.
This matters because medieval drawings are not photographs. A copied diagram may change scale, orientation and curvature while preserving connectivity. A teaching diagram may preserve a relationship while sacrificing physical proportion.
But even a topological match does not establish identity. Blood vessels, rivers, roots, family trees and transport networks all branch. Topology narrows possibilities. It does not name the object.
Blind visual recurrence
The current eduKate visual lane deliberately asks a smaller question than interpretation: do recurring non-semantic visual component families exist across selected pages above matched-control recurrence?
The test uses anonymised visual crops, neutral geometric codes, fixed definitions, matched negative controls, independent or fresh-session scoring, held-out material and a failure ledger. Before scoring, semantic labels such as plant, root, flower, medicine, ingredient, machine part, inventory or Padua are prohibited.
The permitted vocabulary names geometry: axial form, branch event, terminal form, lobe form, band or collar, node or joint, tube or channel, container edge, texture field, symmetry event, attachment geometry and negative-space joint.
The maximum authorised conclusion after the early gate is correspondingly narrow: a candidate neutral recurring visual family has been recovered. Visual recurrence is an existence result. Semantic identity is another experiment.
Agreement between observers is itself measurable
Cohen’s kappa adjusts observed agreement against agreement expected by chance: κ = (pₒ − pₑ)/(1 − pₑ). Yet even kappa does not rescue a vague category. Two scorers can consistently apply the same bad definition. Reliability is not validity.
A strong visual test therefore needs a neutral operational definition, independent application, matched negative controls, a predeclared threshold, replication on unseen material and a record of failed candidates.
Related: The Recurring Visual Grammar Problem: Components Before Objects · The Geometry Tools Problem.
The Time and Materials Warehouse
Radiocarbon dating measures decay, not authorship
Radiocarbon dating is founded on radioactive decay: N(t) = N₀e−λt. The remaining proportion of carbon-14 constrains the time since the once-living material stopped exchanging carbon with the environment.
But the measured quantity does not arrive with an interpretation attached. A parchment date refers to the animal material. A manuscript claim concerns a human production event. The interval between them might be short. It is not guaranteed to be zero.
Science gives a probability distribution over a specified event, not a magical timestamp for the whole object.
Every material has a provenance clock
Parchment, ink, pigment and binding have different analytical possibilities and may belong to different intervention phases. A compatible ink does not prove simultaneous writing across every page. A common pigment does not prove one painter. A later binding does not redetermine the age of the leaves.
Science advances by keeping these evidence clocks separate until a justified relation connects them.
Destructive and non-destructive knowledge
A rare manuscript presents an optimisation problem. Some precise tests require sampling; sampling may consume irreplaceable material. The scientific question is therefore not merely what could be measured, but what we are authorised to damage in order to know it.
Research value − material harm − future-option loss is a useful way to think about the decision. A sample consumed today cannot be tested by a superior method invented tomorrow. A responsible programme preserves future optionality.
The Codicology Warehouse: A Book Is a Physical Machine
Modern readers experience a book as a sequence of numbered pages. A medieval codex is built from folded sheets. A bifolio forms two leaves and four page surfaces. Several bifolios can be nested into a quire. Quires are sewn and bound. Foldouts introduce additional geometries. Leaves can be removed, misplaced, trimmed, rebound or numbered later.
The manuscript’s present order is therefore an observation. Its original order is a historical hypothesis.
Before reading the sequence, prove the sequence.
Codicology becomes graph theory
A manuscript can be modelled as a graph. Pages are connected through different physical relations: same bifolio, same quire, physical adjacency, conjectured original adjacency, shared foldout, missing conjugate, current binding and later foliation. These edges have different strengths and meanings.
A page may be physically joined to one leaf, visually similar to another and distributionally close to a third. That is why a multilayer graph is more faithful than a single narrative sequence.
Missing data are not empty data
A missing leaf is a structured absence. A stub may tell us something was removed; quire structure may constrain where it belonged. But the content remains unknown.
Known absence ≠ known missing content.
Related: Missing Leaves and Quire Reconstruction · The Original-Order Problem.
The Human-Hand Warehouse
Handwriting is a motor trace
Writing records human movement. Stroke direction, curvature, pen lift, spacing, proportions, loop formation and terminal gestures can create a motor signature. Digital paleography turns these features into comparable observations.
Lisa Fagin Davis’s paleographic work proposed multiple scribal hands, commonly represented as a five-scribe model. This creates a richer production hypothesis than the traditional picture of one isolated author. It remains a classification of handwriting behaviour, not identification of five named people, five authors, five languages or a particular city.
A hand is not necessarily an author. A scribe may copy another person’s text. One person may write in more than one style. Different tools, speeds, page positions or time periods may alter the same person’s hand.
Five hands are not a place
Even if five distinct scribes were established beyond dispute, geography would remain unresolved. A collaborative manuscript could have been made in a university environment, a court, a medical or pharmaceutical workshop, a religious house, a private scholarly household, a commercial scriptorium or a distributed copying network.
This is especially important for the Voynich-and-Padua hypothesis. Padua possessed strong mathematical, medical, astronomical, optical and natural-philosophical capabilities. Those capabilities make it a valuable historical environment to study. They show what kinds of knowledge systems and professional interactions were possible. They do not establish that Beinecke MS 408 was produced there.
Capability explains possibility. Only discriminating evidence can establish origin.
Related: Biagio Pelacani: Mathematics, Optics and Natural Philosophy · What the Padua Hypothesis Actually Says.
The Living-Forms Warehouse
Botanical resemblance is not species identification
Many Voynich pages resemble herbal manuscripts. That resemblance is legitimate evidence for comparison. It is not a botanical determination.
A scientific species identification normally requires several independent traits. Voynich illustrations frequently appear stylised, composite, exaggerated or structurally difficult to reconcile with a single living specimen. They might depict real plants imperfectly, copy degraded exemplars, combine features, emphasise diagnostically important parts, represent processed materials, or function as mnemonic or operational diagrams.
The better first question is not “which species is this?” but “which visual features are consistently preserved, and what representation system would make preserving those features useful?”
A drawing is a lossy compression
Every drawing selects. A botanical illustration may preserve leaf shape but omit scale. An anatomical diagram may exaggerate a pathway. A map may distort distance to preserve connectivity. A circuit diagram ignores physical appearance while preserving logical relations.
Reality → selection → encoding → drawing.
The scientific questions become: what information was retained, what was discarded, what was exaggerated, what conventions does the intended reader need, and what task would the representation support?
A drawing can survive while its reading convention is lost. We may possess the representation while lacking the receiver model that once made it useful.
What it looks like ≠ what it does for its reader.
The Astronomy, Optics and Cycles Warehouse
Circular diagrams are among civilisation’s most overloaded forms. A circle may encode celestial motion, calendars, zodiacal divisions, seasonal cycles, cosmology, medical correspondences, elemental relations, logic, geography, ritual order, mnemonic architecture or simply graphic organisation.
A scientific analysis can begin with measurable features: radial divisions, angular regularity, nesting depth, rotational symmetry, repeated labels, directionality, central-versus-peripheral emphasis and relationships among neighbouring diagrams.
A calendar hypothesis should predict a coherent number and arrangement of divisions. An astronomical hypothesis should predict relationships compatible with a specific celestial system. A process diagram should predict transitions or repeated component roles. Geometry converts resemblance into testable expectations.
Mathematics asks what must remain invariant
Suppose two circular diagrams differ in size and artistic execution but share the same number of sectors, nesting order and label placement. That may indicate a common diagram family. Suppose sector count varies freely while the same central-peripheral relation remains. Then the invariant lies at a different level.
Mathematics asks under which transformations the proposed identity should survive. Without invariants, resemblance expands without limit. Anything can be made to look like anything else.
The Padua investigation is useful here as a capability field. Historical figures such as Biagio Pelacani show that questions about vision, geometry, natural causes and astronomical calculation belonged to a real intellectual environment. That context helps us ask better questions; it does not give the manuscript an address.
The Probability Warehouse: Bayes Without Romance
Voynich research contains many hypotheses: natural language in an invented script, ciphered language, verbose or homophonic cipher, constructed language, notation system, formulaic reference system, mnemonic text, generated pseudo-text, deliberate hoax, mixed system—or an explanation not yet represented in the candidate set.
Bayesian reasoning gives us a disciplined way to update them: P(H|E) ∝ P(E|H)P(H). The posterior depends on prior plausibility and on how likely the observed evidence would be if the hypothesis were true.
Much Voynich evidence is shared evidence. Low entropy might fit several language, cipher and generator models. Multiple hands might fit collaboration, copying or changing production conditions. Visual sectioning might fit a reference work, a compilation or a deliberately scholarly imitation.
A hypothesis gains strength by surviving its best alternative
Weak research asks what evidence agrees with a theory. Strong research asks what evidence should exist under that theory but not under its strongest competitor.
If natural language and generated pseudo-text both predict a Zipf-like curve, finding the curve does little to distinguish them. If only one predicts a particular combination of edge dependence, vocabulary tail, boundary behaviour and quire-level stability, that joint pattern becomes more informative.
The best next experiment is therefore the one expected to reduce uncertainty among competing models, not necessarily the one attached to the most spectacular page.
Absence can be evidence—but only under coverage
A missing expected feature counts against a theory only when the search had a fair chance to find it. Negative evidence requires an expected observation, adequate coverage and reliable detection. Without those conditions, absence is an evidence-coverage problem.
Unknown is not none.
The Experimental Science Warehouse
The test that refused to pass
One of the most scientifically valuable results in the current eduKate programme was not a success. A pre-registered text-state challenge required complete exact-locus coverage across 45 validation pages. The locked data aligned on 44. Across the two transcription payloads, 871 of 872 required loci were shared.
The measured association on the aligned pages remained directionally positive. Several numerical gates were satisfied. But one required locus on folio 116v was absent from one locked transcription. The coverage gate required 100 per cent. It failed.
The final test remained sealed. No model was fitted to the final material. No hypothesis was promoted. The failed challenge was preserved as a coverage failure rather than rewritten as a successful effect test.
This is what preregistration is for. Before seeing the result, the programme had declared that incomplete coverage would block promotion. After seeing a promising result, it did not grant itself permission to lower the standard.
Why 871 out of 872 was not “close enough”
The missing locus may expose a systematic difference between data sources; it may belong to a special interface class; it may show that the declared corpus was never fully acquired. The point is not that one missing locus destroys every observation. It does not. The aligned-page effect remains a usable, non-promotional residue.
Interesting residue ≠ passed experiment.
Failed challenger is not permission to retune
After a test fails, several temptations appear: drop the troublesome page, lower the threshold, change segmentation, swap in a corrected transcription, choose another interval, modify the effect definition. Each change may be reasonable for a new experiment. None belongs inside the already observed one.
Corrected data enter as a new versioned challenge. They do not travel backward in time and repair the failed test.
The defeat ledger
A failed theory can leave useful residue. The claim “these pages are an inventory” might fail because quantities, parties, transactions and stock fields are absent. A smaller structural observation may survive: the pages exhibit repeated entry-like geometry. The failed semantic claim is not retained as truth. The structural residue is not discarded with it.
Theory is disposable. Evidence is conserved.
Discovery and validation must be separated
A pattern found and measured on the same material is vulnerable to overfitting. Discovery material suggests the candidate. Validation material tests the frozen candidate. Held-out material asks whether the result travels beyond both.
An explanation that has seen every answer is not yet a prediction.
The Control Warehouse: Known Worlds Before Unknown Worlds
An unknown object cannot calibrate itself. A language hypothesis needs known languages. A cipher hypothesis needs ciphers with known plaintext and construction rules. A pseudo-text hypothesis needs generated samples whose lack of message is known. A botanical interpretation needs identified herbals and image traditions. A catalogue theory needs catalogues.
A comparator is not an ancestor
If a known herbal resembles the Voynich, it is tempting to infer copying. If a Paduan diagram resembles a circular page, it is tempting to infer origin. If a pharmaceutical manuscript shares container-like forms, it is tempting to infer function.
But resemblance may arise from common genre constraints, broad period conventions, convergent solutions, distant common sources or modern selection bias. A comparator may establish that a representation was possible, that a format existed, that a feature was common or rare, or that a method can distinguish known classes. It cannot establish direct contact without a separate transmission chain.
Comparison ≠ provenance. Shared capability ≠ shared workshop.
Hostile controls are better than friendly examples
A friendly comparator resembles the theory. A hostile control could fool it. A visual classifier that claims to recognise medicinal roots should face decorative foliage, river networks, anatomical vessels and invented branch diagrams. A text metric that claims to detect language should face ciphers, formulaic lists and generated pseudo-text.
The best controls are selected to destroy shallow recognition. This is why known medical, astronomical, cosmographic, alchemical and computational manuscripts are so important inside the wider Voynich Research Library.
The Uncertainty Warehouse
Voynich debates often compress every uncertainty into one word: unknown. That loses important structure.
- Measurement uncertainty: was the stroke recorded correctly? Is the gap real?
- Sampling uncertainty: would the pattern persist across other pages, hands or quires?
- Model uncertainty: would another segmentation or control corpus change the result?
- Semantic uncertainty: even if the structure is real, what does it represent?
- Historical uncertainty: when and where did production, transfer or intervention occur?
- Identity uncertainty: are two similar components members of one family?
- Causal uncertainty: did one tradition influence another?
- Completion uncertainty: is the surviving codex a final object, unfinished compilation or damaged remnant?
These uncertainties require different repairs. More pages may reduce sampling uncertainty. Better images may reduce measurement uncertainty. No amount of additional counting automatically resolves semantics. An archive may improve provenance while leaving language untouched.
Precision is not the same as truth
Computers produce exact outputs. The manuscript does not therefore become exact. A script can report an odds ratio to six decimal places, but the meaningful question is whether the event classes were valid, independently reproducible and robust to alternative segmentation.
Many decimal places ≠ strong evidence.
Numerical precision is a property of representation. Epistemic confidence is a property of warranted belief.
The AI Warehouse: An Adversarial Object for Artificial Intelligence
Large language models are exceptionally good at completing patterns. The Voynich Manuscript is exceptionally good at rewarding completion with illusion.
Give a model an unidentified glyph and a possible meaning, and it can construct a fluent explanatory system around it. Give it a city, profession, plant or historical figure, and it can connect the manuscript to that seed through hundreds of plausible associations. The prose may become smoother as the evidence becomes weaker.
This makes the Voynich Manuscript a valuable AI test. Can the system distinguish observation from interpretation? Can it preserve uncertainty? Can it prevent a suggested hypothesis from contaminating later visual scoring? Can it keep text, image, material, historical and provenance lanes separate? Can it record failed predictions? Can it decline to translate an effect size?
AI is useful when its role is bounded
AI can compare transcriptions, calculate distributions, identify candidate anomalies, organise literature, generate hostile alternatives, audit overclaiming, test code, preserve ledgers and propose the next discriminating experiment.
But the model that generates a hypothesis should not automatically be the sole judge of its success. The system that has seen the answer key should not perform the blind scoring. Generator, scorer, validator, evidence custodian and release authority are different roles.
Fluency is not a measurement
An AI answer may be persuasive because it is linguistically coherent. That property belongs to the answer. It does not belong to the manuscript.
A real decipherment must survive stable mapping, independent replication, full-page reading, consistent grammar, explanation of rare forms, illustration-text predictions, scribal and regional robustness, historical plausibility and successful reading of held-out material.
A translation that changes its rules whenever the page changes is not a translation. It is improvisation.
The Grand Inventory: What Did the Mathematics and Science Actually Teach Us?
- The unit must be earned. A glyph is not automatically a letter; a blank-delimited string is not automatically a word.
- Counting needs a contract. Every count requires a unit definition, sample boundary, uncertainty policy and data lineage.
- Information is not semantics. Entropy, mutual information and networks can reveal organisation without revealing meaning.
- Predictability has more than one cause. Low conditional entropy can arise from several generative mechanisms.
- One familiar statistic rarely decides the problem. A theory must explain a discriminating joint profile.
- Boundaries may carry state. A state effect is not a dictionary entry.
- Association is not identity. Variables may interact while remaining distinct.
- The manuscript has overlapping structures. Physical, scribal, visual, textual and distributional provinces do not collapse into one map.
- Current order is not necessarily original order. Codicology must precede narrative interpretation.
- Material dates are event-specific. Radiocarbon dates parchment, not automatically the act of inscription.
- A hand is not an author. Paleography can recover variation without naming people or places.
- Visual recurrence is not semantic function. Blind geometry can establish recurrence without naming the object.
- A comparator is not provenance. Known manuscripts calibrate methods and establish capabilities.
- Failed tests are scientific assets. They can leave smaller reusable residue.
- Frozen rules protect us from ourselves. A threshold changed after observation belongs to a new experiment.
- Held-out material turns explanation into prediction.
- Unknown is not none. Missing coverage is not evidence of absence.
- The best next question separates alternatives.
- AI needs epistemic architecture. Generation, blind scoring, evidence custody and release authority must remain distinct.
- Stopping is part of knowing. A serious investigator must know where the evidence requires the sentence to end.
What We Know, What We Suspect and What We Do Not Know
What is materially established
The Voynich Manuscript is a real historical parchment codex, now catalogued as Beinecke MS 408. Its surviving construction includes foldouts, missing material and complex quire structure. Sampled parchment belongs within an early-fifteenth-century radiocarbon range; scientific examination is consistent with historical rather than modern material production.
What is structurally strong
The script is highly organised rather than uniformly random. Symbol placement is constrained. Apparent token forms have structured distributions. Page regions differ. Boundary position matters. Physical, visual, scribal and distributional groupings recur. Several scribal hands have been proposed.
What remains provisional
The functional status of glyphs, multi-symbol units, apparent tokens and spaces remains unresolved. Local text-state effects require further validation. Visual component families require blind-control and held-out gates. Operational interpretations of page layouts remain hypotheses. Historical capability environments, including Padua, remain comparative contexts rather than proven provenance.
What has not been established
- No plaintext translation is claimed here.
- No language has been securely identified.
- No author has been securely identified.
- No unique purpose has been established.
- No recurrent q-type form is authorised as a fixed translation.
- No conventional visual section name has been proved as semantic content.
- No direct Padua attribution has been established.
- No recurring visual component is licensed here as a medicinal ingredient, machine part, commercial object or named species.
The manuscript survives as an unsolved object. The investigation survives as a working science.
The Return Path: Why the Voynich Manuscript Matters Beyond the Voynich Manuscript
The manuscript’s greatest contribution may ultimately be methodological. Modern civilisation increasingly lives among systems whose outputs are measurable before their meanings are secure: a medical model produces a risk score, an AI system produces an explanation, a financial network produces correlations, a satellite produces an anomaly, a student assessment produces a mark.
In each case, the temptation is the same: convert structure into story before the causal and semantic bridge has been built.
The Voynich Manuscript teaches us to ask: what was directly observed? How was it encoded? Which unit was assumed? Which comparison world was used? What else could have generated the pattern? Which data were held out? What would destroy the claim? Where does the uncertainty live? What returned from the world after the claim was used?
The mathematics of restraint
Mathematics is often presented as a machine for obtaining certainty. The Voynich Manuscript reveals another mathematics: the mathematics of bounded uncertainty. Entropy tells us how much surprise remains. Mutual information tells us where dependency lives. Graphs tell us that there are several kinds of neighbourhood. Matrices tell us that classifications can overlap. Bayesian reasoning tells us that evidence changes hypotheses unevenly. Held-out tests tell us whether a rule travels. Stopping rules prevent desire from rewriting the experiment.
This mathematics does not hand us a translation. It tells us what any translation would have to survive. That is not a lesser achievement. It is the construction of the measuring room in which a future answer could become trustworthy.
The science of not pretending
Science is not the performance of certainty. It is the controlled movement from observation to explanation. We can date material without dating every event; observe recurring geometry without naming objects; detect statistical organisation without identifying language; reconstruct a historical knowledge environment without assigning provenance; find an effect and still refuse to translate it; obtain a promising result and still fail an experiment because one frozen gate did not pass.
That refusal is not a lack of imagination. It is imagination under discipline.
Voynich | The Mathematics and Science
The most interesting result is not that mathematics and science have failed to read the manuscript. They have succeeded in teaching us how an unread object should be approached.
Mathematics gave us the language of frequency, entropy, state, topology, association, networks, uncertainty and prediction. Science gave us material analysis, codicology, paleography, comparative controls, blind scoring, preregistration, failure preservation and the ethics of limited intervention.
Together they produced something rarer than another key.
They produced a way of knowing that does not require pretending.
The Voynich Manuscript remains silent. But around that silence, a remarkably exact architecture has begun to appear. We know which walls are physical. We know which divisions overlap. We know which clocks cannot be merged. We know which patterns recur. We know which explanations reproduce only the easy features. We know which claims have failed. We know what the next experiment is permitted to ask. And we know, with increasing precision, what we still do not know.
That is not the end of the inquiry.
That is the point at which the inquiry finally becomes worthy of its object.
Return to the Voynich Research Library → · Read the Voynich Master Synthesis →