Voynich Decipherment: What a Real Solution Must Prove

The Voynich Manuscript has been solved many times.

It remains unsolved.

That contradiction tells us something important about the problem.

The difficult part is not producing a reading. Human beings are exceptionally good at producing readings from incomplete information. Give us an unknown symbol and we can assign a sound. Give us a plant picture and we can assign a name. Give us a repeated token and we can call it “water”, “root”, “star”, “woman”, “take”, “mix” or “medicine”.

Give a method enough freedom—several possible sound values, omitted vowels, abbreviations, multiple languages, anagrams, null signs, spelling variation and image-led guesses—and it becomes surprisingly easy to produce something that sounds meaningful.

That is not the miracle.

The miracle would be finding a reading that becomes less flexible as it succeeds.

A reading whose rules stop moving. A reading that survives pages the solver did not use while inventing it. A reading that tells us what we should find before we look. A reading that another researcher can reproduce without borrowing the inventor’s intuition.

A real decipherment is not a story that can explain the manuscript after seeing it. It is a constrained model that keeps working when the manuscript is allowed to answer back.

Everything in this series has been moving toward that standard. We separated the physical book from its legend, transcription from translation, statistical clusters from semantic names, image resemblance from provenance, later annotations from original text, and missing pages from imagined content.

Now comes the harder question: if somebody genuinely solves the Voynich Manuscript, what should that solution be able to survive?

Quick Read

One-sentence answer: a genuine Voynich decipherment must specify exactly what text it is reading, use stable and reproducible rules, explain substantial continuous material rather than a few convenient labels, predict unseen passages and visual relationships, account for major structural features such as Currier variation and line-position effects, remain compatible with the physical and historical manuscript, outperform realistic alternatives, expose its failures, and be reproducible by independent researchers.

A plausible translation is not automatically a decipherment.
The input must be reproducible: images, transcription, glyph segmentation, space policy and treatment of uncertain characters.
EVA characters are transliteration labels, not known sounds.
Rules should become harder to change as the solution develops.
Rules created from showcase passages should be fixed before validation elsewhere.
A serious solution must read substantial continuous text, not merely isolated names or labels.
It should work on pages not used to build the proposed reading.
Repeated forms should behave consistently unless a principled linguistic or encoding rule explains variation.
The proposal needs an explicit mechanism for how visible writing becomes the proposed underlying information.
Currier A/B and other major text variation cannot simply be ignored by a whole-manuscript solution.
Line beginnings, line endings and paragraph openings need explanation.
Labels, running text, circular text and Quire 20 records should be handled according to their distinct document roles.
Images must provide independent validation rather than supplying the answer that they later “confirm”.
The method must respect bifolia, missing leaves, uncertain page order and later marginalia.
The proposed language, notation or encoding method must be historically plausible for the early-fifteenth-century material world.
The mechanism should explain the manuscript’s unusual statistical properties as well as produce meaningful output.
Realistic alternative models matter; weak alternatives make favourite theories look artificially strong.
Independent replication is essential for broad acceptance.
Exceptions and failures must remain visible.
A partial decipherment should be called partial. “Solved” should be the last word earned, not the first word advertised.

What Does “Solved” Actually Mean?

The word is used too broadly. Identifying one plant is not the same as reading the writing system. Finding one plausible proper name is not the same as recovering a grammar. Producing one meaningful paragraph is not the same as explaining thousands of repeated forms across several textual regimes.

A useful ladder is:

Observation: a pattern exists.
Interpretation: one explanation fits the pattern.
Local reading: one form or short passage receives a plausible reading.
Partial decipherment: stable rules explain a bounded class of text and predict new examples.
System decipherment: a coherent model explains substantial regions of the writing system.
Full functional interpretation: text, images, physical architecture and historical use become mutually coherent.

A valid partial decipherment would be an extraordinary achievement. It does not need to be called a full solution to matter. Claim size should match evidence size.

The Input Must Be Reproducible

Before a proposed solution tells us what a Voynich word means, it needs to tell us what characters it thinks the word contains.

A serious method should identify the image source, transliteration alphabet, transcription file, segmentation rules, treatment of ambiguous forms, treatment of rare characters and policy for visible spaces. Circular and radial writing also require an explicit ordering convention because a modern transcription has to choose where a circle becomes a line.

If a translation depends on one disputed stroke being read one particular way, the final conclusion inherits that uncertainty. A certain modern sentence cannot legitimately emerge from an uncertain medieval glyph without the uncertainty travelling through the argument.

Every decipherment has an input layer. If the input cannot be reproduced, neither can the solution.

The Rules Must Freeze

Exploration begins with alternatives. A glyph might represent one value, another value or part of a compound. That flexibility is normal while a hypothesis is being invented.

Validation is different. As evidence accumulates, a successful theory should progressively lose freedom. If one glyph was assigned a value to read Page A, it should not quietly become something unrelated on Page B merely because the second value produces a nicer word.

Context-sensitive behaviour is completely possible. Languages change forms by context. Historical abbreviations vary. Encoded systems may use multiple states. But the context rule itself must be explicit and should predict when the change occurs.

A theory must become harder to change as it explains more.

If success creates more discretionary freedom rather than less, the method is probably fitting the examples rather than recovering the system.

A Showcase Passage Is Not Enough

The manuscript contains thousands of word-like units. One impressive label or sentence explains a tiny fraction of the problem.

Short strings are especially easy to fit when the method allows abbreviation, multiple sound values, omitted vowels, spelling variation, several languages or image-led candidate meanings. Long continuous text is less forgiving. Grammar has to persist. Frequent forms reappear. Morphology recurs. Discourse needs structure.

Coverage therefore matters. The larger the uninterrupted passage read under fixed rules, the harder it becomes for coincidence and discretionary interpretation to carry the result.

Read Text You Did Not Use to Build the Key

Suppose a researcher studies ten herbal pages and develops a mapping that turns them into plausible medicinal prose.

The next step is not to celebrate those ten pages.

Freeze the mapping. Choose pages that were not used while developing it. Do not change the character values. Do not add a new exception merely because the hidden page is difficult.

If the system remains coherent, confidence rises. If it collapses, the original pages may have been overfit.

The pages used to invent a key are not the strongest pages on which to validate it.

This is ordinary out-of-sample reasoning applied to a medieval manuscript.

Prediction Is Stronger Than Recognition

Imagine two cases.

In the first, a plant looks like mint, so the solver translates the nearest label as a historical word for mint. Image and text now agree—but the image supplied the answer.

In the second, a textual rule is derived elsewhere, frozen, and applied to an unseen plant label. The output predicts a plant identity or property before the image is used. The image can now validate independently.

The second event is rarer and much stronger.

After-the-fact resemblance is abundant. Correct prediction is scarce. A real solution should generate scarce events.

Frequent Forms Cannot Mean Whatever the Picture Needs

Highly frequent Voynich tokens create a severe consistency test. If one form occurs hundreds of times, a proposed reading has to explain why that same underlying form appears in all those contexts.

Natural language allows multiple meanings. Grammar changes word forms. Encoded systems can use complex mappings. None of that requires a naive one-token-one-English-word model.

But variation needs rules. A common token cannot become “plant” on an herbal page, “woman” in Quire 13 and “star” on a zodiac page simply because those meanings suit the illustrations. If the meanings differ, the theory has to explain the linguistic or encoding mechanism that permits the difference.

Frequency turns unconstrained flexibility into a liability.

There Must Be a Mechanism, Not Only a Gloss

A list of translated words is not yet a model of a writing system.

A genuine proposal should explain how the visible script produces the proposed underlying information. Is the system alphabetic, syllabic, abbreviation-heavy, coded, enciphered, constructed, or hybrid? What are the transformations? What remains stable? What changes by context?

Each mechanism class predicts different behaviour. An alphabetic orthography predicts reusable sound–form relationships. A code system predicts recurring semantic groups. A verbose encoding predicts expansion between underlying and visible units. An abbreviation system predicts historically plausible contractions and expansions.

“This token means rose” is a gloss. “This visible sequence is generated from underlying form X by rule Y, which also predicts these other forms” begins to become a decipherment mechanism.

Grammar Must Emerge From the Source

Natural-language solutions face an additional test. Long text should behave grammatically.

The grammar need not look like modern English. A medieval language can have unfamiliar word order, spelling and morphology. But nouns, verbs, function words and inflections still follow constraints.

If every line becomes grammatical only after the translator inserts unstated verbs, changes word order, supplies omitted subjects and chooses unrelated senses for common tokens, the target language may be carrying more structure than the manuscript.

The grammar should be recovered from repeatable source-side relationships rather than donated by the translation.

The Exception Budget Must Be Visible

Historical manuscripts contain errors, spelling variants, abbreviations and difficult passages. A decipherment is allowed exceptions.

The problem begins when exceptions are uncounted and infinitely expandable. One glyph changes value here. One word becomes an anagram there. One line switches language. Another needs omitted letters. A failure is called a scribal error. Any individual move may be historically possible; hundreds of discretionary moves create a system capable of explaining almost anything.

A strong proposal should report how many special cases it needs, what kinds they are, and whether the rules for those cases existed before the difficult examples appeared.

A solution should not become stronger by hiding its patch count.

The Governing Principle

A method flexible enough to explain every outcome has not yet explained the Voynich Manuscript. A real solution should reduce freedom, increase prediction and leave the manuscript more capable of proving the solver wrong.

Currier A and B Are Not Optional

A whole-manuscript solution has to confront one of the most durable textual facts: the Voynich text changes state. Currier A and Currier B differ in character sequences and word-like distributions, and later analyses continue to find large-scale non-uniformity even when exact classification schemes differ.

If A and B are two dialects, a solution should show the systematic linguistic relationship. If they are two orthographies, show the transformation. If they are two encoding states, show what changes state. If they reflect different source traditions, writers or topics, demonstrate the independent variable rather than simply renaming the clusters.

Both regimes occur inside broadly plant-dominated material, which already makes the easiest “A means plants, B means bodies” interpretation too simple. Quire 13 is strongly B-like, but B also occurs elsewhere. A whole-book theory cannot use Currier labels as semantic conclusions while leaving their cause unexplained.

A cluster is an observation about distribution. A decipherment should tell us what historical or linguistic process generated the cluster.

Line Position Is Not Optional Either

Voynich writing behaves differently at line beginnings and endings. Paragraph openings have distinctive properties. Some glyph families disproportionately occupy special positions.

A proposed natural language can explain positional effects through syntax, discourse and scribal conventions. An encoding system can build position into its rules. A generative system can do the same. What a full solution cannot do is translate the visible words while treating the positional structure as somebody else’s problem.

If the plaintext reads smoothly but the proposed mechanism gives no reason why visible forms change at line boundaries, the semantic layer has not yet explained the writing system. A translation can be meaningful and still be generatively incomplete.

Paragraph Openings Create Another Validation Layer

Paragraph beginnings often carry discourse information. A recipe may begin with a formula equivalent to “take”. A herbal entry may begin with a name. A catalogue record may begin with an identifier. If distinctive Voynich forms prefer paragraph openings, a genre or language theory should predict their role.

Conversely, if the opening forms are decorative or scribal rather than lexical, their behaviour should reveal that. Paragraph position is therefore not merely another statistic. It is a place where competing mechanisms can make different predictions.

A whole solution should tell us why beginnings look like beginnings.

Labels Are a Separate Challenge

Labels are among the easiest units to overinterpret because they sit beside visible objects. A short string beside a plant fragment looks like a plant name. A string beside a zodiac figure looks like a personal or star name. A string beside a vessel looks like its contents.

But labels can encode names, categories, numbers, status, cross-references, abbreviations or short instructions. A decipherment that claims “these are names” should show name-like behaviour across many labels and should distinguish labels from ordinary running text under a principled grammatical or documentary model.

A label solution that cannot leave the label environment may still be a valid partial decipherment. The claim should simply remain bounded.

Circular and Radial Text Must Not Inherit a Modern Start Point

Some diagram text runs around circles or along radii. A modern transliteration must turn that geometry into a linear character stream. That operation creates choices: where does the circle begin, which direction does it run, and is there an original first token at all?

If a decipherment assigns syntactic importance to the first transcribed token in a ring, it must show that the manuscript itself marks the same start. Otherwise the grammar may be responding to a modern editorial cut rather than medieval structure.

This is one of the cleanest tests of whether a proposed method is reading Beinecke MS 408 or reading the conventions of the transcription file.

Quire 20 Gives Hundreds of Opportunities to Be Right

The final star-marked gathering is one of the manuscript’s best validation environments because 324 marginal star-like markers survive, most associated with short textual units. The missing central bifolium means the original total was probably larger, but the surviving corpus still offers hundreds of candidate records.

If these are recipes, recurring textual roles should emerge: ingredient-like forms, quantity-like forms, operation-like forms, formulaic openings and perhaps repeated administration language. If they are registry records, a different field structure should emerge. If they are summaries or cross-references, rare labels from earlier sections may recur disproportionately.

A genuine decipherment should not need to display only the five friendliest entries. The record architecture gives it hundreds of chances to demonstrate that its grammar generalises.

Missing Pages Cannot Repair a Failing Solution

Fourteen folios are missing under the old foliation. A central bifolium is absent from Quire 20. F74 is cut out near a visually interesting transition. These losses genuinely increase uncertainty.

They do not provide positive evidence for imagined content. If a theory needs a missing key, missing table, missing zodiac diagram or missing transition to work, the solver may say that such a page could once have existed. The solver cannot count the imagined page as support.

Unknown pages are uncertainty, not a repair kit.

Images Must Be Independent Witnesses

The pictures are both a gift and a major source of circularity. If a plant picture causes the solver to expect “mint”, and the nearest token is then adjusted until it becomes a word for mint, the picture cannot subsequently be presented as independent validation.

The stronger sequence reverses the dependency. Derive enough of the textual system without the target image. Freeze the rules. Apply them to an unseen label. Then reveal the picture. If the textual output correctly predicts an independently recognisable object or property, the image has become a witness rather than a prompt.

The same standard applies beyond plants. A textual rule might predict a zodiac category, a vessel class, a Quire 13 state or a relationship on the Rosettes foldout. Cross-modal prediction is unusually valuable because the evidence streams can disagree.

The Herbal Pages Should Not Become a Self-Proving Dictionary

The easiest circular workflow is: identify the picture, assign the nearby token the plant name, then use that token as proof that the identification was correct.

A stronger plant-based partial decipherment would need several independently diagnostic plant identifications, one stable textual mapping, historically plausible names or synonyms, and successful prediction of plant identities on pages not used to build the mapping.

One spectacular plant match is weaker than a systematic set of ordinary predictions.

The Rosettes Foldout Is a Valuable Hostile Case

Every broad theory has favourite pages. A herbal theory loves the plant pages. An astrological theory loves the zodiac. A hydraulic theory loves Quire 13. A geographical theory loves Rosettes.

The pages a theory dislikes may be the ones that matter most for validation. A whole-book solution must eventually explain why the Rosettes foldout contains nine connected major regions, local labels and extensive writing. It cannot simply designate the foldout “symbolic” because its preferred story works better elsewhere.

Difficult pages are not annoyances. They are opportunities for the manuscript to distinguish a local fit from a global system.

Vessel Shapes Cannot Become Free Semantic Wildcards

If a proposed solution says that vessel morphology encodes preparation methods, the visual classes should predict text. Similar vessels should correlate with similar operation language; different classes should generate systematic differences.

If the vessels instead encode storage or categories, another pattern should appear. What matters is not which interpretation sounds most plausible. What matters is whether the decoded text makes a prediction about the visual classification that the solver did not use to create the translation.

Earlier attractive “jar shape = process” ideas remain hypotheses until a stable cross-page legend is demonstrated.

The Physical Book Must Agree With the Semantic Book

A decipherment happens inside a codex. If a proposed semantic sequence requires two leaves to be consecutive, codicology should not strongly contradict the relationship. If a theory depends on an exact entry total, missing material must be accounted for. If it assumes a modern section boundary, the physical gathering structure should be checked.

Bifolia matter because one original sheet can connect leaves separated in modern reading order. Quire structure matters because production batches may correspond to scribal or textual states. Later binding and numbering matter because present order is not identical by definition to original reading or production order.

Semantics cannot overrule parchment.

Chronology Is a Hard Constraint

The sampled parchment belongs to the early fifteenth century, commonly reported within 1404–1438 at 95% probability. A proposed writing system, language, technical vocabulary or encoding method therefore needs a historically plausible route into that world.

A technique documented only centuries later is not automatically impossible, but its historical burden rises. A plant identification with major post-Columbian implications is not automatically impossible, but its botanical and transmission evidence must be correspondingly strong.

Historical plausibility is part of the mechanism, not decorative context added after the translation has already been accepted.

Historical Plausibility Is Not Provenance

A hand-operable fifteenth-century-compatible mechanism can be plausible without being proven to have been used. A language can exist in northern Italy without proving Padua. The Carrara Herbal can demonstrate sophisticated medicinal illustration near 1400 without becoming an ownership record for Beinecke MS 408.

A decipherment may eventually generate new provenance evidence through dialect, names or technical vocabulary. That would be strongest if geography emerges after the language model has been derived rather than being assumed beforehand and then “confirmed” by words selected from that region.

Statistics Are Not a Side Quest

Voynichese has measurable structural properties: unusual character predictability, word-family behaviour, line-position effects, local vocabulary, Currier variation and long-range distributional organisation. A proposed mechanism should explain why these arise.

A fluent underlying translation is not enough if the proposed encoding would generate surface statistics unlike the manuscript. The reverse is also true: a mechanism can reproduce Voynich-like statistics while carrying no demonstrated meaning.

The strongest outcome joins the layers: one historically plausible mechanism both generates the observed surface structure and recovers coherent underlying information.

The 2013 Language-Like Results Set a Floor, Not a Finish Line

Studies published in PLOS ONE in 2013 found important non-random organisation in Voynich text. Long-range token distributions and network properties differ from shuffled text and are compatible in important respects with natural-language corpora.

These results make naive randomness an inadequate explanation. They do not identify the underlying language or mechanism. A real solution should explain what generates the long-range organisation—topic, source, language, encoding state or some combination—rather than stopping at the word “language-like”.

The 2025 Naibbe Result Raises the Standard for Rejecting Ciphertext

Michael Greshko’s 2025 Cryptologia study showed that a historically plausible hand-operable verbose homophonic substitution system can produce several unusual Voynich-like statistical properties from meaningful Latin or Italian plaintext.

This does not solve Voynich and does not prove that the Naibbe system was used. Its methodological importance is that it strengthens the alternative model. A solver can no longer rely on the simple argument “language-like statistics therefore not ciphertext”.

Natural-language hypotheses should compete against realistic encoded alternatives. Encoded hypotheses should compete against realistic language and structured-generation alternatives. The strongest model is the one that explains evidence on which strong competitors make different predictions.

An Encoded-Text Solution Must Explain More Than Substitution

A simple one-for-one substitution is attractive because the rule is easy to state. But simple substitutions preserve many underlying frequency relationships, while Voynich has unusual character-level predictability and token structure.

More elaborate historical encoding hypotheses can introduce homophones, verbose groups, nulls, multiple tables or context-dependent states. Those mechanisms may fit more observations, but every added mechanism increases flexibility. The complexity therefore has to earn itself by predicting unseen visible forms and by recovering coherent underlying information.

“The system must be complicated because the page is difficult” is not a mechanism. A real model states which complications exist, when they apply, and what new observations follow from them.

A Natural-Language Solution Must Explain the Strange Surface

Suppose Voynichese is a natural language written directly in an unusual script. The solution still needs to explain why character sequences are so constrained, why many word-like forms differ by small changes, why some glyphs strongly prefer certain positions, why line beginnings differ from endings, why Currier regimes exist and why vocabulary is so local.

Natural languages can create all of these phenomena in principle. But “natural language” is a category, not a finished causal explanation. The proposed orthography, morphology and grammar must generate the observations rather than merely coexist with them.

A fluent translation that leaves the visible system unexplained is only half a decipherment.

A Structured-Generation Theory Must Pay the Semantic Bill

Suppose the text was generated by a constrained process without ordinary plaintext. Such a process can potentially explain word families, low entropy, positional forms and local repetition.

Then a different burden appears. Why was the system generated? Why does it change across page regions? Why is it integrated with labels, plants, zodiac rings, human figures and record markers? Why did a maker invest so much material and labour in the object?

A meaningful-plaintext theory must explain the strange surface. A meaningless-generation theory must explain the organised manuscript around the surface. Neither gets the difficult half for free.

f116v Is Not a Free Bilingual Key

The final verso contains mixed writing that has been read in several ordinary-language ways, with Voynich-script forms on the same page. This makes it an irresistible candidate for a key.

But two scripts sharing a page are not automatically translations of one another. A real bilingual or biscriptal key requires alignment: evidence that corresponding segments express related content, stable mappings between repeated elements, and successful prediction elsewhere in the manuscript.

No accepted bridge currently turns f116v into such a key. It remains high-value historical evidence and low-certainty decipherment evidence.

The Later Month Names Are Not a Free Language Key Either

The zodiac month names are readable Romance-language additions in a different hand. They show that a later reader understood the central diagrams in conventional zodiac/month terms and can help reconstruct reception.

They do not establish that the main text uses the annotator’s language. A later French- or Romance-writing person can annotate an older manuscript created elsewhere and written in another system.

A proposed solution may use the month names as historical context. It cannot treat spatial proximity on the parchment as proof of linguistic identity.

External Anchors Are Valuable Because They Remove Freedom

Undeciphered scripts often become tractable when researchers find strong external anchors: a bilingual text, known ruler, predictable date formula, repeated title or securely known astronomical value.

Voynich has weaker candidate anchors: zodiac emblems, later month names, plant-like images, star-marked record boundaries, mixed f116v writing and known provenance points. These are useful, but most do not align source and target language as directly as a true bilingual inscription would.

A strong solution uses anchors in proportion to their independence. A recognisable zodiac ram is a strong iconographic anchor. The meaning of the thirty surrounding figures is not equally anchored. One should not inherit the certainty of the other.

A Solution Should Help Explain Why Historical Readers Failed

By the seventeenth century, Georg Baresch was already seeking help with the book, and Marci sent it to Kircher because its writing was not understood. The knowledge gap is therefore old.

A real solution may reveal why: perhaps the notation depended on a lost professional convention, a specialised abbreviation system, an encoding procedure, a rare language variety or a key no longer transmitted.

Historical failure is not disproof of a simple solution, but it becomes another datum. If the proposed system would have been immediately readable to an educated seventeenth-century reader, the theory should explain why the manuscript was already being treated as mysterious.

Independent Replication Is the Moment a Private Insight Becomes Public Knowledge

A researcher may discover something through intuition, unusual tools, private exploration or years of accumulated experience. Discovery does not need to be tidy.

Public validation does. Another researcher should be able to receive the stated input, rules and exception policy, apply the method and obtain comparable outputs without the inventor standing beside them making discretionary decisions.

If only the inventor can tell when a glyph changes value, which language applies, whether a word is abbreviated or which image resemblance matters, too much of the method remains private intuition.

A solution becomes knowledge when it can leave its inventor.

Replication Should Not Depend on One Favourite Transcription

A proposed decipherment may legitimately require a better segmentation than basic EVA. The visible writing system may not align perfectly with any existing alphabet.

But if a custom transcription is essential, the custom distinctions should be demonstrated on the parchment. Another researcher should be able to identify the same graphical classes without knowing the desired plaintext first.

If a solution collapses under another major transcription, there are two possibilities: the custom representation captures a real distinction better, or the reading depends on convenient segmentation choices. Return-to-image validation should decide between them.

Falsifiability: What Would Make the Solver Say “I Was Wrong”?

This question quickly separates a testable model from a self-sealing one.

If contradictory text can always be called corrupt, a wrong image can always be called symbolic, a failed reading can always switch language, and an inconvenient page can always use another encoding state, evidence cannot defeat the theory.

A serious proposal should state failure conditions. Perhaps the fixed mapping fails on withheld text beyond a stated level. Perhaps proposed grammar repeatedly generates impossible forms. Perhaps plant predictions fail. Perhaps the mechanism cannot reproduce observed statistics. Perhaps independent researchers cannot reproduce the segmentation.

Failure conditions make success informative.

A Theory That Explains Every Outcome Predicts Nothing

Suppose each glyph may have several unrelated values, vowels may be inserted or omitted freely, strings may be read forwards or backwards, multiple languages may be selected by context and inconvenient words may be treated as anagrams or nulls.

Each operation can exist in a real historical system. The problem is the combined freedom. With enough choices, short meaningful outputs become cheap.

The Voynich Manuscript does not need a method capable of finding meaningful words. It needs a method incapable of finding most wrong words.

Constraints are what make a successful reading surprising.

A Good Solution Should Compress the Manuscript

Before a solution, we need thousands of separate observations: this glyph occurs here, this token repeats there, this line has a positional effect, this page is B-like, this label belongs to a star.

A genuine model should explain many observations with fewer principles. One orthographic rule may explain several glyph variants. One grammatical rule may explain a word family. One encoding rule may explain unusual predictability. One genre model may explain Quire 20 record structure.

This is explanatory compression. A system with hundreds of exceptions is not compressing much; it may be rewriting the manuscript in a larger vocabulary of special cases.

Simplicity Helps, but the Simplest Story Is Not Automatically True

A simple substitution may be elegant and wrong. A more complex historically plausible encoding system may be justified if the extra mechanisms explain several otherwise independent properties and make successful predictions.

The goal is not minimum complexity at any cost. It is minimum necessary complexity for the evidence.

Every added mechanism should buy explanatory power across multiple observations, not merely rescue one difficult word. Necessary complexity generalises. Ad hoc complexity patches.

The Best Solution Should Survive Its Worst Pages

Every broad theory has a friendly region. The strongest validation often comes from the unfriendly region.

What does a herbal theory do with Rosettes? What does a city theory do with hundreds of plant pages? What does a straightforward natural-language model do with strong line-position effects? What does a simple encoding theory do with Currier A/B?

A whole-manuscript theory should eventually become more useful on the pages that originally looked inconvenient. Those pages are where the object has the greatest chance of proving that the model was too narrow.

Partial Solutions Are Allowed—Boundaries Must Be Sharp

A real partial decipherment could transform the field. Someone might decode one numerical system, one label class, one Currier regime or one later annotation family before the rest of the manuscript becomes readable.

The scientific problem begins only when the boundary is blurred. “We can explain this bounded label class” becomes “we solved Voynich”. A constrained partial result is easier for later scholarship to build upon than an inflated result that mixes demonstrated and speculative claims.

Precision in claim size is not modesty for its own sake. It is infrastructure for future work.

“It Makes Sense” Is a Weak Validation Standard

A proposed translation can sound wonderfully appropriate: “Take the red root and wash it in warm water.” If the page contains a plant and coloured root, the sentence feels persuasive.

But target-language coherence is cheap when a translator can choose synonyms, insert omitted grammar, smooth word order and ignore difficult forms. The stronger question is not whether the English makes sense. It is whether fixed source-side rules force that reading substantially more strongly than plausible alternatives.

Meaningfulness is necessary. It is not sufficient.

A Long Translation Can Still Be Wrong if the Rules Drift

Length alone does not validate a translation. Pages of coherent paraphrase can be generated if every source unit has several possible values and target-language grammar is supplied flexibly.

Coverage has to be paired with constraint. How many mappings changed? How many exceptions appeared? How many source tokens were left untranslated? How much reordering or unstated grammar was supplied?

A 200-word translation under rigid rules that predicts new passages can be much stronger than a 5,000-word translation whose rules continuously adapt.

One Famous Proper Name Is Not Enough

Short labels can often be made to resemble famous names—cities, physicians, plant genera, stars—especially when historical spelling variation and abbreviation are available.

A strong proper-name identification should constrain the rest of the page. If one label is a city, neighbouring labels should fit the same spatial system. If one is a plant name, the same phonetic or orthographic mapping should work for other independently identifiable plants.

A proper name is powerful when it generates consequences. Isolated resemblance is weak even when the name is exciting.

A Real Model Should Predict Error Patterns Too

Scribes make mistakes. A credible model can sometimes predict which mistakes are likely. Visually similar glyphs may be confused. A repeated encoded group may be copied incorrectly. An abbreviation may be expanded inconsistently. Different writers may favour different variants.

This is much stronger than labelling every failed example a scribal error. Even failure should have a distribution. If “scribal error” appears only when the proposed translation is inconvenient, it is a patch rather than an explanatory component.

Handwriting Models and Decipherment Models Should Eventually Meet

The main-hand question remains live. Lisa Fagin Davis’s influential palaeographic work identifies five hands; a 2026 critique by Torsten Timm argues that the variation may be more continuous.

A decipherment does not need to settle this debate immediately, but it can create new evidence. If one underlying word appears in predictable graphic variants by proposed hand, that supports a scribal-variation model. If A/B differences disappear after controlling for one writing style, the causal picture changes.

Ultimately handwriting, textual regime and semantic system should become mutually informative rather than one disputed classification being treated as unquestioned truth.

Different Textual Jobs Should Produce Organised Differences

Labels, long paragraphs, star-marked records and circular diagram text need not have identical statistics even if they use the same underlying writing system. Genre changes vocabulary and syntax.

A strong solution therefore should not demand uniformity everywhere. It should predict the non-uniformity. Labels may be noun-heavy. Recipe-like records may contain repeated operations. Running exposition may contain more grammatical connectors. Diagram text may use compact naming or relational structures.

Variation becomes evidence when the semantic model predicts why it occurs.

What a Real Decipherment Should Do With the Pictures

Eventually, text and image should constrain one another in both directions. A good textual model might predict that two plant pages discuss related substances, that a detached fragment corresponds to an earlier whole plant, that one zodiac label belongs to a repeated class, that a Quire 13 pool colour marks a particular state, or that one vessel family aligns with one textual category.

These are powerful because they cross modalities. The text says something about the image system before the image is used to choose the words. The image can then agree or disagree independently.

The manuscript’s illustrations are too organised to remain decorative after a genuine whole-book solution. As decipherment improves, the pictures should become less mysterious for principled reasons rather than because every picture is renamed to match the translation.

What a Real Decipherment Should Do With the Physical Book

The physical architecture should also become more intelligible. Why do some textual regimes cluster by bifolium? Why does Quire 13 form such a coherent visual gathering? Why does Quire 20 use repeated star-marked records? Why do foldouts carry particular kinds of diagrams?

A semantic model may reveal that these features correspond to workflow, source batches, scribes, modules or use. It does not need to explain every binding decision immediately. But the strongest eventual account will make text, image and codex converge rather than leaving them as three unrelated mysteries.

What a Real Decipherment Should Do With Provenance

A decipherment can create new provenance evidence. A securely identified dialect can narrow geography. Repeated named places can transform the map. A physician, patron or institution could connect documentary records to the codex. Technical vocabulary can reveal a professional community.

But provenance is strongest when it emerges downstream. If a researcher assumes Padua, selects Paduan vocabulary, and then announces that the vocabulary proves Padua, the conclusion has returned to its starting point.

Independent convergence is strongest when the text arrives somewhere the historical research did not force it to go.

What a Real Decipherment Should Do With Uncertainty

Not every glyph will be clear. Not every page survives. Not every drawing is identifiable. A good solution does not need fake certainty.

It can say that one glyph is ambiguous, one word has two plausible senses, one plant identification is tentative, or one historical localisation remains broad. The key is calibration: weak inputs should not produce certain outputs merely because certainty is rhetorically satisfying.

A method becomes more trustworthy when it can leave some material unresolved rather than changing itself until every token receives a confident gloss.

What a Real Decipherment Should Publish

For independent evaluation, a serious proposal should publish enough of its final method and evidence for another researcher to reconstruct the claim:

the input transcription or exact manuscript loci;
the character or grapheme inventory;
segmentation rules;
mapping rules;
the linguistic, notational or encoding mechanism;
known exceptions;
the passages used to develop the model;
the material held back for validation;
complete outputs, including failures;
historical assumptions;
comparison with plausible alternatives.

This does not mean researchers must publish every private notebook, exploratory branch or proprietary tool that helped them discover the hypothesis. Discovery and validation are different activities.

The public claim becomes assessable when the final method can survive without access to the private route that discovered it.

Discovery Can Be Messy; Validation Cannot Stay Private

Real research is rarely a tidy staircase. A useful idea can begin from an accidental visual resemblance, a computational anomaly, a failed theory or an intuition that is initially impossible to formalise.

None of that weakens the final result. What matters is what happens next. Once the hypothesis exists, the public validation should be expressible independently of the private route that produced it.

The world does not need every private thought. It needs enough evidence and method to test the public claim.

This preserves both legitimate research privacy and scientific accountability.

Seven Common Pseudo-Solution Patterns

These are general failure shapes rather than accusations against any particular researcher.

1. The One-Label Solution

One short label resembles a historical name, but the mapping cannot read sustained text.

2. The Picture-First Dictionary

The picture suggests a meaning, the token receives that meaning, and the same picture is then presented as proof.

3. The Moving Alphabet

Character values change whenever the current value fails to produce a convenient word, without a stable context rule.

4. The Unlimited-Abbreviation Solution

Any missing letters can be supplied because historical scribes abbreviated, leaving so many possible expansions that success becomes cheap.

5. The Multilingual Escape Hatch

A token is Latin when Latin fits, German when German fits and Italian when Italian fits, without a rule predicting the language switch.

6. The Missing-Page Rescue

Whenever a required key or transition is absent, it is placed on a lost folio and treated as evidence.

7. The Meaningful-Paraphrase Solution

The final modern-language passage sounds coherent, but the route from source units to target meaning contains too much unreported discretion.

Any one operation used in these patterns can be legitimate. The failure occurs when flexibility is not constrained enough for the manuscript to say no.

A Real Solution Should Make New Research Easier

A genuine decipherment should do more than close an old mystery. It should open new research that other fields can use.

If plant names become readable, botanists can reassess the images. If the text is medical, historians can compare remedies and practice. If a language variety is identified, linguists can investigate its geography and morphology. If an encoding mechanism is recovered, historians of cryptology can reconstruct workflow. If Quire 20 is truly a recipe corpus, historians can compare its formulae with known manuscripts.

Truth creates interfaces. A private interpretation that only its inventor can use tends to isolate the manuscript further.

What Would Convince Us That a Plant Label Had Been Deciphered?

A strong case might look like this:

The glyph sequence is stable under a published segmentation.
The same mapping reads several other labels.
The candidate names belong to one historically plausible language or nomenclatural system.
The plant identifications are based on independent morphology rather than on the proposed text.
The mapping predicts the name or property of an unseen plant page.
The prediction matches the picture independently.
The same rules continue into other text environments or the partial boundary is clearly stated.

At that point, the field may genuinely have a bounded decipherment rather than an attractive label match.

What Would Convince Us That Quire 20 Contained Recipes?

A strong recipe reading should recover repeated textual roles across many star-marked records: formulaic openings, ingredients, quantities, operations, administration instructions and perhaps indications or conditions.

Ingredient-like forms should cross-reference independently identified plant or pharmaceutical labels. Quantity-like forms should occupy constrained positions. Operation-like terms should recur across different ingredients. The grammar should predict records that were not used to develop it.

Only then does “recipe section” move from a useful modern nickname toward demonstrated genre.

What Would Convince Us That Rosettes Was a Map?

A map reading should recover map-like information: stable place labels, spatial relationships, routes, boundaries, landmarks or orientation conventions. Identifying one architecture-like form is not enough.

If one region is identified as a particular place and another as a second place, the proposed geography should predict what lies between them. The connector network should make sense under the same model, and labels not used in the original identification should become readable.

A map is a system of spatial constraints. A real map solution should recover the system, not merely a photogenic landmark.

What Would Convince Us That Quire 13 Was About Baths?

Historical therapeutic-bath literature makes balneology plausible. A decipherment could transform that plausibility into identification if recurring labels resolve into attested bath or water terms, long text consistently describes appropriate regimens or conditions, pool classes correlate with different vocabulary, and connector structures receive stable functions.

The same rules should predict an unseen Quire 13 page. Until then, the existence of medieval bath medicine demonstrates a domain the manuscript could occupy; it does not prove that it does.

The Solution Should Not Depend on Knowing the Answer First

If applying a method requires already knowing that the page is herbal, that the plant is aloe, that the language is Italian and that the adjacent word should mean aloe, the method is formalising an expectation rather than recovering an answer.

A stronger method can be handed to somebody who does not know what the target page depicts. It produces a reading under fixed rules. Then the image is revealed and either agrees or does not.

That is the moment the manuscript begins to validate itself rather than reflecting the solver’s prior belief.

Independent Convergence Is the Strongest Destination

Imagine textual decoding, statistics, images, codicology and history all developed with partly independent evidence and then converging on one model.

The text predicts a place; documentary history independently supports the region. Plant labels predict species or substances; morphology independently agrees. A production model predicts bifolium-level differences; palaeography independently finds them. A recipe grammar predicts recurring ingredient labels; the pharmaceutical pages independently supply the same labels.

World-class confirmation is different evidence streams arriving at the same place without being told in advance where to go.

What the Retained Work Changed Most

The most useful outcome of examining many branches was not discovering one irresistible total explanation. It was becoming less impressed by explanations that survived mainly because they could change whenever the manuscript objected.

Several ideas were historically plausible. Some matched individual pages beautifully. Some created elegant whole-book stories. When the explanatory freedom was reduced and the entire object was allowed to push back, their certainty weakened.

The reader-safe result is what remained: structure is real; local organisation is real; visual families are real; physical constraints are real; several attractive semantic bridges remain unproven; and the best future solution will need to be more predictive than the stories that came before it.

That raises the floor for the next serious claim.

The Reader’s 20-Point Decipherment Audit

Input: What exact images or transcription are being read?
Alphabet: What counts as one glyph or grapheme?
Spaces: Are visible spaces assumed to be lexical words?
Rules: Were values fixed before validation?
Coverage: How much continuous text is read?
Held-out test: Does the method work on unseen material?
Consistency: Do common forms behave under stable rules?
Grammar: Does source-side structure generate coherent language rather than free paraphrase?
Mechanism: Is the writing system explicitly modelled?
Exceptions: How many special cases are required?
Currier: Does the model explain major textual regimes?
Position: Does it explain line and paragraph effects?
Text roles: Are labels, paragraphs, circular text and records treated appropriately?
Images: Are pictures independent validators or sources of candidate meanings?
Codicology: Does the interpretation respect bifolia, missing leaves and uncertain order?
History: Could the proposed system exist in the relevant period?
Statistics: Can the mechanism explain major Voynich structural properties?
Alternatives: Has it competed with realistic other models?
Replication: Can independent researchers reproduce it?
Falsification: What observation would make the theory fail?

A proposal does not need perfection on all twenty points to deserve attention. The more ambitious the claim, however, the more of the list it must survive.

Primary School: A Rule Must Keep Being the Same Rule

Give a child a simple code: triangle means A, circle means B, square means C. Show triangle–circle–square and read ABC.

Then show another message where ABC is inconvenient and ask whether triangle may suddenly mean T because T produces a nicer word.

No—unless the code already contains a rule explaining when triangle changes. The child has just learned the foundation of decipherment: the rule must survive the answer.

Lower Secondary: Build on One Set, Test on Another

Give students ten coded messages. Let them use five to infer the system. Then reveal the other five and forbid rule changes.

One team may look brilliant on the first five but fail the hidden five. Another may produce a less exciting initial theory that generalises. The second team has the stronger model.

This teaches held-out validation without advanced mathematics.

Upper Secondary: Count Degrees of Freedom

Compare two toy decoding methods. Method A gives every symbol one fixed value. Method B gives each symbol four values, allows inserted vowels and permits words to be read backwards.

Both produce one meaningful phrase. Which success is more surprising? Method A, because it had fewer ways to succeed accidentally.

This is the intuitive meaning of degrees of freedom. A flexible theory needs stronger validation than a constrained one.

JC and Adult Readers: Think in Out-of-Sample Prediction

At a higher level, decipherment is a model-selection problem. Several models can explain observations already seen. Generalisation is what distinguishes them.

Voynich offers many out-of-sample targets: unseen folios, a different visual section, rare labels, Currier regime changes, star-marked records and cross-modal visual predictions.

A model that performs across these domains has captured something deeper than local coincidence.

A Parent and Teacher Guide

The Voynich problem can teach a complete reasoning discipline without requiring children to solve the manuscript.

Define the input. What exactly are we reading?
Keep rules stable. Do not change them merely to obtain a nicer answer.
Separate discovery from proof. A clue can inspire a theory without validating it.
Test something unseen. Prediction is stronger than explanation after the fact.
Count exceptions. Every patch should remain visible.
Use independent evidence. One clue should not create and prove the same answer.
Respect missing data. Unknown pages cannot be filled with convenient evidence.
Compare alternatives. A theory gains strength when serious competitors fail where it succeeds.
Ask what would change your mind. A theory that cannot lose cannot truly win.
Match the claim to the evidence. A partial success can remain partial and still be excellent work.

These habits belong to cryptanalysis, but also to science, history, journalism, statistics and everyday decision-making. The manuscript becomes a classroom for learning how to deserve certainty.

Frequently Asked Questions

Has the Voynich Manuscript been deciphered?

No proposed decipherment has achieved broad scholarly acceptance across the manuscript. Many local and global readings have been proposed.

What is the strongest single validation idea?

No single test is sufficient, but one of the strongest is whether fixed rules developed on one set of material successfully read or predict substantial unseen material without discretionary changes.

Is a meaningful sentence enough?

No. Short strings can often be made meaningful under flexible mappings. A decipherment needs reproducible rules, broad coverage and successful prediction.

Does a solution need to translate every word?

No. Historical manuscripts contain damage, rare terms and uncertainty. A partial decipherment can be valid if its boundaries are explicit and its rules predict new examples.

Must a repeated Voynich token always have one modern meaning?

No. Natural language has polysemy and morphology, and encoded systems can create complex mappings. Variation must follow a principled rule rather than convenience.

Does a natural-language solution have to explain statistics?

A strong whole-manuscript solution should. Its orthography and grammar should help explain observed character, token, positional and regional distributions.

Does a statistical model count as decipherment?

Not by itself. Reproducing Voynich-like statistics is evidence about mechanism. A semantic decipherment additionally needs recoverable underlying information.

What does the Naibbe study show?

It shows that a historically plausible hand-operable encoding system can reproduce several unusual Voynich-like statistical properties from meaningful plaintext. It does not prove that the Voynich Manuscript uses that system.

Can the pictures prove a translation?

They can provide powerful independent support when the textual reading was not invented from the same image. If the picture supplied the candidate meaning, agreement is not independent validation.

Is f116v a bilingual key?

No accepted evidence establishes that. A true key would need stable alignment between the ordinary-language-looking writing and Voynich-script forms and would need to predict material elsewhere.

Do the later month names reveal the main language?

No. They belong to a later hand and are stronger evidence about a later reader than about the principal text.

Does a real solution need independent replication?

Yes for broad scholarly acceptance. Other researchers should be able to apply the stated method and obtain comparable results without relying on the inventor’s private intuition.

What is the safest current conclusion?

The manuscript contains a highly structured writing system whose physical, textual and visual constraints are strong enough to make a genuine decipherment testable, even though no accepted solution has yet passed the full set of constraints.

Research and Further Reading

The Final Idea

The Voynich Manuscript does not need another clever reading.

It needs a reading that can survive being made less clever.

Take away the freedom to change character values. Take away the freedom to choose only friendly pages. Take away the freedom to let pictures supply words and then prove those same words. Take away the freedom to place missing evidence on missing leaves. Take away the freedom to call every failure a scribal error or switch languages whenever a word refuses to cooperate.

What remains?

If the answer is still a coherent system; if it reads material it has never seen; if repeated forms behave consistently; if Currier A and B make sense; if positional effects make sense; if the pictures begin confirming things the text predicted; if the physical book agrees; and if another researcher can reproduce it—then something remarkable has happened.

The manuscript has stopped being a mirror for the solver’s imagination.

It has started constraining the answer.

A real decipherment is not the explanation we can keep alive. It is the explanation the manuscript cannot easily kill.

Voynich | Everything eduKate Knows and Tested | What a Real Voynich Decipherment Must Survive