A message can become longer without saying anything new. Sometimes that is poor writing. Sometimes it is how the message survives.
Imagine that someone tells you a meeting is on Tuesday, then adds the date, and finally names the day again while saying goodbye. The information overlaps. If you mishear one part, another part may rescue it. But suppose the speaker has copied the wrong date from a calendar and repeats that same error three times. Repetition has made the statement more insistent, not more reliable.
That difference gives us a sharper way to approach the Voynich Manuscript. Rather than asking only why forms recur, we can ask whether any recurrence gives a reader another route to the same information. Does the writing contain constraints that could expose an error? Could a damaged sign be reconstructed because other signs restrict it? Would a historical reader have recognised a mistake that remains invisible to us?
Those are questions about redundancy. They are not yet evidence that Voynich contains a deliberately engineered error-correcting code. The distinction matters because an unread manuscript invites two opposite mistakes: treating every repetition as meaningless padding, or treating every regularity as proof of a hidden protective design.
The central question is not whether Voynich repeats. It is whether something useful remains recoverable because of what repeats.
Return to the Voynich Research Library for the wider evidence map. This article follows The Noise Problem, but asks a different question: not which differences are harmless, but what protects a distinction when part of its representation is lost.
What can we currently say?
Voynich’s recurring forms and constrained symbol sequences make redundancy a reasonable subject of investigation, but they do not establish deliberate error protection. The cited public research does not recover a historical decoding rule that corrects Voynich errors. A successful redundancy claim would need to identify what information is repeated, which disturbances the repetition can resist, and how a reader could recover the intended state without inventing it.
Yale continues to describe the text as undeciphered and its script as unidentified. That is the status from which this discussion begins, rather than a claim that an error-correction system has been discovered. The worked examples below are constructed illustrations, not measurements taken from the manuscript and not reports of a newly executed Voynich experiment. [1]
There are three claims to keep separate. A sequence may be predictable. Its constraints may make some damaged sequences recoverable. Its maker may have intended that recoverability. Each claim is stronger than the one before it, and each needs additional support. We can investigate the first two without pretending to know the third.
Read through the distinctions, a worked example, the manuscript evidence, competing explanations, recovery without translation, historical intention, and the reader workshop.
Three distinctions that prevent most mistakes
Repetition is an observation; redundancy is a relationship
When the same visible string appears twice, we have observed recurrence. We have not established that the two occurrences convey the same message. A common word can occur in two different instructions. A shared ending can perform the same grammatical job on different nouns. A mark can identify a class rather than repeat an entire proposition.
Redundancy concerns overlap in information relative to a specified task. Suppose a record writes an amount once in figures and once in words. The two expressions look different, yet they may constrain the same intended quantity. Conversely, two identical labels on different pages might refer to different local items if each page uses its own numbering system. Visual sameness and informational overlap are therefore not interchangeable.
For Voynich, the responsible first question is: redundant about what? A glyph identity, a word form, a grammatical relationship, an entry boundary, an object’s identity, or a whole instruction? Without an answer, the word redundancy can conceal a change of subject.
Predictability is not automatically error correction
A highly predictable sequence may be easy to continue without carrying a recoverable message. Consider a page consisting only of alternating circles and squares. You can fill a blank correctly, but that success tells you little about whether the original page communicated instructions, names or observations. You have learned its pattern, not necessarily its content.
Now consider a system that can express several distinct messages, but represents each through a constrained pattern. If damage occurs, the receiver might use those constraints to distinguish the intended message from its alternatives. That is a different achievement. There is something to recover, a defined set of candidates and a rule for choosing among them.
The Voynich question lies between these possibilities. A model that restores a missing glyph may be exploiting spelling-like regularity, a generation template or a genuine encoding constraint. The restoration alone does not identify which explanation is correct.
A useful effect is not proof of an intended design
Suppose a writing convention turns out to resist some copying errors. It does not follow that a historical author invented the convention for that purpose. The protection may be a side effect of grammar, familiar formulae or a restricted vocabulary. A reader could benefit from a constraint that no scribe ever described as error control.
Deliberate protection is therefore a historical claim. It would become more plausible with evidence of checking practices, systematic repair, parallel notation, explicit conventions or a relevant documented tradition. Without such evidence, the stronger formulation is that a structure is consistent with error tolerance, not that its creator engineered a code.
These distinctions protect the research from anachronism. Modern information theory can help us ask precise questions about an old manuscript without requiring its makers to have possessed modern mathematical concepts.
A small example: when repeating three times really helps
Here is a deliberately simple system. There are only two possible messages, represented by 0 and 1. Instead of writing one mark, the sender writes the chosen mark three times. Message 0 becomes 000. Message 1 becomes 111. The reader uses a majority rule.
If 000 is damaged into 010, the majority still points to 0. If 111 becomes 101, the majority still points to 1. The extra marks do not introduce another message choice. They provide a margin of protection around the original choice. This is an illustration of a repetition code, not a proposed Voynich alphabet.
Assume, only for this example, that each mark independently has a one-in-ten chance of flipping. A one-mark message is then wrong ten percent of the time. A three-mark majority is wrong only when two or three positions flip. Its error probability is 3 × 0.1² × 0.9 + 0.1³, which is 0.028, or 2.8 percent.
The calculation tells us exactly why the redundancy helps. It also tells us its cost: we have used three written positions to carry one binary message choice. The result is not free reliability. It is a trade between the amount written and the probability of recovering the intended message under a particular error model.
The broader mathematical distinction between source information, redundancy and protection against a noisy channel is foundational to Shannon’s communication theory. What matters for this article is that protection depends on how redundancy is arranged and on what can go wrong, rather than simply on the number of repeated marks. [2]
Now change the failure, and the benefit can disappear
Suppose all three copies are produced from the same misread original. The writer reads 0 as 1 and faithfully writes 111. Nothing in the majority rule detects the mistake. All three copies agree because they share the same error.
Or suppose damage reverses the interpretation of the whole group together. If that event has a ten-percent probability, repeating three times does not produce the earlier 2.8-percent result. The independent-error assumption has failed. The apparent backups are not independent backups.
This is especially important when thinking about manuscript transmission. Repeated text inherited from one exemplar may preserve an ancestral error with extraordinary consistency. Agreement can demonstrate common descent rather than correctness. A claim that Voynich protects its own information must therefore name the kinds of disturbance against which protection is supposed to work.
Detection and correction are different achievements
A system may tell us that something is wrong without telling us how to fix it. Imagine writing three binary message positions and adding a fourth position so that the total number of 1s is even. If one position flips, the total becomes odd. The reader can detect a problem, but the evenness rule alone does not reveal which position changed.
For example, 0000 could become 0001, 0010, 0100 or 1000. An odd total signals trouble in each case. It does not, by itself, supply a unique original message. Two changes can restore an even total and escape that simple check altogether.
Applied to Voynich, a proposed grammatical or structural rule might identify an unusual form as suspicious. That is not permission to rewrite it. The rule may detect a deviation while leaving several plausible restorations. A rigorous reading must preserve that ambiguity rather than replacing it with the most convenient word.
What the public Voynich evidence actually supplies
Lindemann and Bowern’s comparative study found unusually constrained character placement in Voynich transcriptions and examined how glyph composition, writing conventions and other representation choices affect entropy. The result makes local predictability a substantive feature to explain; it does not identify an error-correcting system or recover a source message. [3]
That observation is a starting point, not a destination. The existing entropy article asks where predictability comes from. The present article asks what that predictability would allow a receiver to reconstruct if the written evidence were partly obscured. A strong explanation should connect those questions without treating them as identical.
Reddy and Knight’s public analysis also treats word forms, sequence structure and position as separate objects of investigation. That separation is useful here: regularity inside a transcribed token is not automatically regularity across a line, and neither is automatically redundancy of meaning. [4]
A predictable ending may be useful without repeating a word
Suppose a certain token family strongly restricts its final component. A reader who recognises the beginning might reconstruct an obscured ending. No complete word needs to be repeated elsewhere for that recovery to become possible. The overlap lies between the visible beginning, the family’s construction rules and the missing position.
But the restoration might recover only the expected spelling pattern. If several legitimate meanings share that pattern, the semantic question remains open. We should distinguish recovering an expected glyph sequence from recovering what the sequence meant to its intended user.
This is why the target must be specified before any apparent success is counted. A glyph-restoration task, a word-identity task and an instruction-recovery task may all use the same page, yet they test different levels of knowledge.
Dense word families can work against easy correction
The manuscript’s familiar near-neighbour forms create a useful complication. Suppose changing one component takes one valid token into another valid token. The alteration may no longer look like an error. It looks like a legitimate alternative.
Our toy 000/111 system avoided that problem because the two valid messages differed in three positions. Their nearest one-change variants could be assigned to the intended original. A family in which valid alternatives sit very close together offers less protection at that local level unless some wider context distinguishes them.
The inference is conditional, not a verdict against redundancy. Close visible forms could be equivalent spellings, cipher alternatives or different meaningful words. Until those relations are known, we cannot count every near-match as either damage or protection. The Word Families article remains the place to examine their distribution.
A compound can create apparent redundancy in transcription
Imagine that a scribe treats one complex mark as a single sign, but a transcription represents it as two characters. The first transcribed character may then predict the second almost perfectly. That dependency is useful for restoring the transcription. It does not demonstrate that the historical writer added a second independent check.
Matlach, Janečková and Dostál investigated whether some complex Voynich symbols could be compounds or ligatures. Their proposed decomposition is a hypothesis, not an agreed decipherment, but the work illustrates why the chosen analytical units affect claims about the system’s structure. [5]
Before calling one component redundant, ask whether it is actually an optional repetition, a necessary part of one sign, or an artefact of the way the sign has been written in a modern file. Those alternatives imply different historical mechanisms.
Local regularity should not be promoted into whole-book protection
A model might restore line endings well and still fail elsewhere. That would establish something about line endings, not about the resilience of the manuscript’s entire content. A model might also work within one textual regime but fail when vocabulary or writing conventions change.
The appropriate response is to preserve the scope. Successful local reconstruction is valuable. Its value is reduced, not increased, when it becomes a claim that an unknown language has been recovered. Careful boundaries let small findings accumulate without forcing them into a premature solution.
Six explanations that can produce overlapping information
Grammar can repeat relationships
In an ordinary language, more than one part of an expression can signal a grammatical relationship. A repeated cue can help a reader interpret a damaged sentence without being a separately designed checksum. The exact conventions depend on the language, and no Voynich language has been recovered here.
A linguistic explanation would become more persuasive if the allegedly redundant cues formed stable grammatical relationships across many contexts. It would become weaker if every repeated component needed a new grammatical description. The task is to discover a reusable relation, not to give every recurrence a plausible linguistic name.
For that reason, grammatical redundancy should be investigated alongside the Function-Word Problem, not used as a substitute for it. The existence of repeated-looking markers does not establish their grammatical functions.
Document templates can make entries predictable
Imagine a catalogue whose entries always provide a name, a class and a location in that order. Even without reading the names, a user can notice that one entry is incomplete. The template supplies expectations. This kind of reliability belongs to document organisation rather than to a cryptographic code.
A hypothetical Voynich entry system could behave similarly. Repeated openings, endings or separators might make missing components noticeable. Yet an organised layout cannot tell us whether the omitted component was a quantity, an object name or a procedural instruction. The template establishes a slot before it establishes the slot’s meaning.
The distinction also explains why a highly formulaic page might be easy to complete superficially. Restoring a familiar pattern is not necessarily restoring the particular information that once occupied the pattern.
Cipher alternatives can add variation rather than backup
Suppose one underlying unit can be written in several legal ways. The alternatives may conceal repeated source material rather than repeat it visibly. Their existence increases the number of possible written forms, but it does not automatically increase the reader’s protection against error.
One can imagine a cipher with carefully separated alternatives that tolerate damage. One can also imagine alternatives so similar that small alterations create confusion. Both are possible mechanisms. The word homophony does not settle the question.
This article therefore does not treat cipher multiplicity as redundancy by default. A claimed benefit must be demonstrated at the layer where an intended receiver recovers information. Extra choice for the writer is not the same thing as an extra independent clue for the reader.
Copying can preserve patterns without preserving truth
A copying process naturally creates recurrence. If a scribe reproduces a phrase several times, the resulting agreement may reflect one source rather than several independent confirmations. If the exemplar is wrong, faithful copying preserves the mistake.
The distinction matters even for modern analysis. Two transcriptions are not independent witnesses merely because they have different filenames. One may have been derived from the other. A consensus assembled from dependent readings can look stronger than its underlying evidence.
None of this tells us that Voynich was copied in a particular way. It tells us what a redundancy argument must establish before counting repeated representations as separate evidence. Their history matters as much as their agreement.
A generator can make missing forms easy to predict
A constrained generator may produce strings from a small number of templates. If its rules are learned, some missing pieces will be recoverable. That does not make the generated output a meaningful source language. It shows that regular generation can create recoverable surface structure.
This is why the existing Self-Citation Problem remains relevant as a competing explanation. A model must distinguish information protection from the simple fact that related forms are produced repeatedly.
A recovery demonstration becomes more informative when equally repetitive but differently generated material is considered too. The important comparison is not between Voynich and unstructured random marks alone. It is between several systems that can all produce regularity for different reasons.
Text and image can overlap—or supply different information
If a hypothetical label names exactly what an illustration depicts, the two channels may partly overlap. If the illustration identifies the object while the text gives a quantity or action, their information is complementary instead. Both arrangements can be useful, but only the first is straightforward redundancy about identity.
The difference becomes important when one channel is obscured. A picture might help recover the identity of an object while providing no clue to an amount written beside it. Calling the whole page redundant would hide the fact that different components protect different things.
For Voynich, the relationship between a label and its neighbouring image remains a question to establish, not a dictionary to assume. A plant-like illustration cannot validate a guessed plant name when the same illustration supplied the guess.
What can be recovered before anything is translated?
A useful research target need not be a word meaning. One might ask whether a withheld, otherwise legible mark can be reconstructed from context. That is a question with a known answer because the original reading has been temporarily hidden, not because the manuscript has been deciphered.
This distinction is essential. Artificially hiding a clear mark creates a controlled restoration problem. Trying to reconstruct an actually lost mark creates an historical inference problem whose answer may be unavailable. Success in the first can inform the second, but it cannot turn an uncheckable restoration into a direct observation.
Any published claim should therefore state whether it concerns exact visible-form restoration, recovery of an equivalence class or recovery of meaning. A model that correctly predicts a common ending should receive credit for that result and no more.
A blank and a wrong mark are not the same problem
When a position is blanked deliberately, the model knows where uncertainty lies. When one plausible mark is substituted for another, the model must first notice that anything is wrong. In a text with many similar-looking legal forms, the second task may be much harder.
An inserted or missing mark adds another complication: the boundaries may shift. A restorer who assumes that every later position remains aligned can misinterpret the consequences. In manuscript work, smudging, omission and mis-segmentation therefore deserve separate treatment rather than one undifferentiated label of noise.
The implication is practical. A claim that a system tolerates ten percent damage has little meaning until the damage is described. Ten percent known blank positions, ten percent plausible substitutions and one continuous missing region are not equivalent challenges.
Recovering the commonest form is a weak victory
Suppose most hidden endings belong to one very frequent form. A model that always chooses that form can perform well without understanding anything about context. The apparent recovery rate should therefore be compared with what a deliberately simple guess would achieve.
It is also important to separate rare forms from common ones. A system might restore familiar patterns accurately while failing on the names or quantities that would carry the most distinctive content in a hypothetical technical book. An impressive average can conceal precisely the cases in which information matters most.
The most useful result is not merely that some marks can be guessed. It is that specified contextual evidence improves recovery of specified distinctions beyond straightforward frequency-based expectations.
Neighbouring examples can leak the answer
If a model learns from most of a page and is tested on a few omitted tokens from that same page, it may benefit from local repetition. That can be appropriate for a narrowly defined restoration task. It is weaker evidence for a mechanism intended to generalise across the manuscript.
A broader claim needs broader separation between the material used to formulate the rule and the material used to assess it. A result that transfers to a different document role or physical unit answers a more demanding question than one that reconstructs neighbours of examples already encountered.
There is no need to pretend one arrangement is always correct. The test should match the claim. Local recovery can be useful locally; manuscript-wide conclusions need evidence that has actually travelled beyond the local pattern.
A restored pattern should not overwrite the source
Once a model expects a form, it becomes tempting to treat unusual readings as errors and silently normalise them. That can produce a cleaner corpus while removing genuine exceptions. The model then appears to fit better because the evidence has been changed to resemble it.
A careful restoration should remain a separate proposal beside the observed reading. The disagreement is useful information. It might reveal damage, a transcription mistake, an unrecognised sign or a weakness in the model itself.
This restraint is not merely archival etiquette. It keeps the inference reversible. Future researchers must be able to see what the manuscript supplied and what an analytical rule supplied in addition.
What would justify saying the repetition was on purpose?
The title asks about intention, so the answer must eventually return to history. A modern demonstration that a pattern resists noise would establish a capability. It would not establish why the pattern was used.
Stronger evidence could include a documented checking convention, independently understood duplicate expressions, recurrent corrections that restore a shared constraint, or a close historical comparator whose users demonstrably relied on the same arrangement. None of those should be invented to complete a satisfying story.
The inference would become more persuasive if several such lines converged. A visible rule, a practical recovery benefit and evidence of how readers used that benefit would together support a claim about intention more strongly than any one measurement alone.
Few corrections do not prove an error-proof system
An apparently clean text can result from careful copying, preparation elsewhere, a tolerant writing convention, unnoticed mistakes or difficulty recognising corrections in an unknown script. Error protection is only one possible explanation.
There is also a difference between a system that prevents mistakes and a system that makes them recoverable. A code may tolerate an error while leaving the incorrect mark visibly unchanged. Conversely, a careful scribe can produce a clean page using a system with little formal protection.
The existing discussion of correction therefore remains independent evidence. It should be connected to a redundancy model only after the model specifies what kinds of mistake its intended users could notice and how they would respond.
A checksum would need a reproducible rule
A dramatic possibility is that one position checks information elsewhere, rather than adding another ordinary content unit. The responsible way to discuss this is as a hypothesis with a high burden. It would require a defined relation between the supposed check and the material it checks.
The relation should work before a translation is chosen. It should not require changing the arithmetic, moving the boundary or discarding an awkward sign whenever a new line appears. It should also distinguish between legal exceptions and failures rather than absorbing every mismatch.
Nothing in the research cited here licenses naming a particular Voynich glyph as a checksum. The value of the idea is that it specifies what evidence would be needed, not that it supplies a ready-made discovery.
The reader’s knowledge belongs in the explanation
A modern reader can often recognise a misspelled familiar word because vocabulary and context restrict the alternatives. A historical specialist might likewise recover a compressed expression that is opaque to an outsider. In that case, some of the protection comes from the reader’s knowledge rather than from repetition on the page alone.
This connects redundancy with the Mnemonic Problem. A cue can be effective for a trained reader while insufficient for someone who lacks the associated knowledge. But imagined expertise cannot become an unlimited explanation for every gap.
A good account specifies what the reader already knows, what the page adds and what remains recoverable when either source is weakened. Otherwise, a claim about the manuscript’s reliability may quietly be a claim about an invented reader who always knows the answer.
Where the redundancy hypothesis is most likely to fail
One failure would be a recovery rule that works only because the target was represented twice in the analytical input. For instance, a hidden character might still be available indirectly through a duplicated transcription field. That is an input problem, not evidence of historical redundancy.
Another failure would be a rule that repairs ordinary cases by deleting the rare distinctions that matter. A model can increase apparent regularity by collapsing several valid states into one. Its output becomes predictable precisely because it has stopped preserving the differences we wanted to understand.
A third failure would be apparent protection that disappears under a more realistic disturbance. A model robust to isolated blanks might be helpless when an entire line is missing. A model robust within one hand might fail when an unfamiliar shape is encountered. Those limitations should be reported as part of the result, not treated as inconveniences.
Most importantly, successful restoration might still leave semantics unidentified. That would not make the restoration worthless. It would mean that an analytical tool has recovered surface regularity while the meaning question remains open. Reliable research allows that intermediate outcome to stand on its own.
New representation research sharpens the boundary
A 2026 preprint by Rozanova and Temerev argues that visible glyphs, blank-delimited tokens and spaces should not automatically be treated as letters, words and uniform word boundaries. Whatever survives further scrutiny of that proposal, it is relevant to recovery claims: the unit being reconstructed must be defined rather than assumed. [6]
This article does not adopt a recovered alphabet from that preprint. Its narrower consequence is methodological. An analysis should state whether it restores a modern transcription unit, an independently supported compound or a visible mark in the image. Switching between those targets can change the apparent amount of redundancy.
A reader workshop: learn the distinction without touching the manuscript
For younger readers: when two copies disagree
Write a short instruction twice and change one word in one copy. Ask a learner whether the two copies tell them that a mistake exists, and whether they tell them which copy is correct. The answers can differ. Repetition may expose disagreement without resolving it.
Add a third copy prepared independently from the original. The learner now has a possible majority clue. Then prepare all three copies from the same incorrect version. The majority becomes confidently wrong. This simple contrast teaches why independence matters more than the sheer number of agreeing copies.
The lesson does not require a mysterious script. It concerns ordinary evidence: several statements can agree because they are correct, or because they share one source. The question worth asking is how the agreement was produced.
For secondary readers: detection without repair
Use the four-position evenness example described earlier. Begin with a valid pattern, alter one position and ask learners to detect the change using the evenness rule. Then ask them to name the original pattern without seeing it. They will discover why a warning is not a unique repair.
Next, alter two positions. Some corrupted patterns now pass the check. The rule has not become useless; its limits have become visible. A responsible description states the kinds of error it can detect instead of calling the system secure or reliable without qualification.
Bring the lesson back to manuscripts. A rare Voynich sequence may violate a proposed rule. That makes it a useful place to look. It does not automatically tell us what the scribe should have written, or whether the proposed rule is correct.
For advanced readers: specify the receiver
Compare two hypothetical receivers. One knows only the allowed symbol patterns. The other also knows the message vocabulary and the document’s purpose. Give them the same partly obscured text. The second receiver may recover more, but the improvement cannot all be credited to marks preserved on the page.
This exercise clarifies a subtle accounting problem. When we say a manuscript contains enough information for recovery, do we mean by itself, or together with a trained reader’s knowledge? Neither question is illegitimate. They simply require different descriptions of the evidence.
For Voynich, this distinction prevents a modern model’s language knowledge from being mistaken for information directly recovered from the historical object. A plausible completion may reflect a useful prior. It still needs a reason to be accepted as the manuscript’s original content.
What would an earned conclusion look like?
An appropriately modest result might say that a specified class of visible Voynich forms can be restored from surrounding structure better than a defined baseline, under a stated kind of artificial masking. That would be a useful finding about recoverability. It would not establish translation or historical intention.
A stronger result might show that several independent constraints converge on the same restoration, that the rule transfers beyond the material used to formulate it, and that it preserves rare distinctions rather than flattening them away. This would make the structural interpretation harder to replace with a convenient alternative.
A historical conclusion about deliberate protection would need still more: evidence that the relevant conventions belonged to the makers’ or readers’ practice. The progression is therefore observation, demonstrated recovery, explanation of the recovery, and only then an argument about intention.
What does not follow is equally clear. Predictability does not prove meaningfulness. Redundancy does not prove a cipher. A successfully restored transcription does not prove a recovered language. Several agreeing copies do not prove independent evidence. An unknown historical reader does not give us permission to supply unlimited missing knowledge.
Questions worth keeping beside a Voynich claim
What is supposedly repeated? Ask whether the claim concerns visible marks, an underlying unit, a grammatical relation or a meaning. A theory that changes levels during its explanation has not yet defined the information it protects.
What disturbance is supposed to be resisted? Ask whether the example concerns a known blank, a wrong sign, a missing stretch or a shared copying error. A successful answer for one kind of disturbance may say little about another.
What can the receiver recover, and what do they already know? Ask whether the result is a glyph prediction, a possible word, a unique message or a plausible story. Then ask what information entered from outside the manuscript to produce it.
What would make the claim fail? The answer should be more concrete than the discovery of a different theory. It might be poor recovery of rare forms, failure on another text population, dependence on a transcription decision, or inability to separate the model from a repetitive generator.
Sources and scope
[1] Yale Beinecke Library: The Beinecke Cipher (Voynich) Manuscript. Institutional description and access to the digital manuscript. Used here for the manuscript’s undeciphered status, not as evidence of an error-correction mechanism.
[2] Claude E. Shannon: A Mathematical Theory of Communication (1948). The foundational distinction between information, redundancy and noisy communication. The binary examples in this article are illustrative derivations, not historical claims about Voynich.
[3] Luke Lindemann and Claire Bowern: Character Entropy in Modern and Historical Texts: Comparison Metrics for an Undeciphered Manuscript, revised 2021. Comparative analysis of character predictability and representational choices.
[4] Sravana Reddy and Kevin Knight: What We Know About the Voynich Manuscript (2011). Public computational analysis of several levels of textual structure.
[5] Vít Matlach, Barbora Anna Janečková and Daniel Dostál: The Voynich Manuscript: Symbol Roles Revisited (2022). Research on symbol behaviour and possible compounds; its proposed interpretation is not treated here as a recovered alphabet.
[6] Liudmila Rozanova and Alexander Temerev: A Glyph Is Not a Letter, a Token Is Not a Word, a Space Is Not a Space (2026 preprint). Used as a recent argument for explicit unit definitions, not as a decipherment. Sources checked for this article on 5 September 2026.
The final idea: protection has to protect something
The most useful change in perspective is small. Instead of seeing a repeated form and asking what secret it must conceal, ask what would become harder to recover if that form were absent. Sometimes the answer may be nothing. Sometimes a boundary, a pattern or an identity may become less certain.
That difference can be studied without inventing a translation. It can also be stated honestly without promoting every regularity into a medieval communications theory. The manuscript may contain protective overlap, ordinary grammar, habitual repetition, constrained generation or several of these together.
A convincing redundancy argument will not be the one that counts the most repetitions. It will be the one that shows what those repetitions let an appropriately defined reader recover—and what they still do not.
Continue through The Noise Problem, The Key-Recovery Problem, and The Cross-Reference Problem, or return to the Voynich Research Library.