Voynich Longform 03 · MECHANISM

How the Voynich Manuscript Works | The Machine We Cannot Yet Read

Do not translate it yet.

Do not decide whether it is Latin, Italian, Hebrew, a cipher, a hoax, a medical handbook or a private notation system.

For this article, all of those questions are temporarily downstream.

We begin with a different question: what kind of machine would have to exist to produce the observable Voynich text?

This is the third article in the new Voynich sequence. HUMAN examined the receiver. KNOWLEDGE stabilised the factual floor. MECHANISM now watches the system move.

If we cannot yet read the output, we can still study the machinery that constrains the output.

---

1. The machine-first rule

A mechanism-first investigation asks what transformations are visible before deciding what the symbols mean. The input to our analysis is not a translated sentence. It is a sequence of marks, spaces, lines, page locations, illustrations and physical divisions. From these we can measure recurrence, position, neighbourhood, variation and dependence.

The central discipline is to keep three layers separate. First comes surface behaviour: what the manuscript visibly does. Second comes mechanism class: the kinds of procedures that could generate those behaviours. Third comes semantics: what the system may encode or communicate.

Most bad decipherments skip from surface resemblance to semantics. MECHANISM stays in the middle long enough to ask whether the proposed engine can actually reproduce the manuscript.

2. Six mechanism families enter the room

For this article we keep six broad mechanism families alive at once.

• Natural language: the visible tokens represent meaningful linguistic units directly or with limited transformation.
• Abbreviation or compression: writing condenses ordinary language through scribal conventions.
• Cipher or encoding: an underlying message is transformed into a different surface system.
• Template generation: tokens are built from restricted slots or reusable structural patterns.
• Copy-and-modify: new forms are produced by locally copying and changing earlier forms.
• Mixed system: two or more of these mechanisms interact.

These are not mutually exclusive final answers. They are working families. A real manuscript could contain language plus abbreviation, language plus cipher, or several production regimes across sections. The point is to give every observation somewhere to push.

3. Surface marks are the first state

The most primitive observable state is a mark on parchment. A mark has shape, stroke order, position, size, orientation and neighbours. Before we call it a letter, we have already made a classification decision.

This matters because a mechanism can operate below the level of the transcription alphabet. Two forms that analysts transcribe as separate symbols may be variants of one underlying stroke pattern. A complex sign may be a fused sequence. A rare form may be an ornamental or positional variant.

MECHANISM therefore treats the glyph inventory as a model to be tested, not a sacred input. A robust mechanism should not depend completely on one fragile decision about how a difficult mark is tokenised.

4. From marks to glyphs

Once recurrent shapes are grouped into glyph classes, the manuscript becomes a discrete sequence. This allows counting. We can ask which glyphs are common, which pairs occur, where certain forms prefer to stand and which combinations appear to be avoided.

This is already a mechanical description. A glyph that almost always appears in one position behaves like a constrained component. A glyph that substitutes for another within families may participate in a transformation rule. A glyph that appears mainly at paragraph beginnings may belong to layout or scribal procedure rather than ordinary lexical content.

The key is not to name the rule too early. First detect the constraint. Then ask which mechanism families can produce it naturally.

5. Spaces create word-like units

Visible spaces divide the text into token-like units. Researchers usually call them words because that is operationally useful. But MECHANISM uses a more cautious term: spaced tokens.

If the spaces correspond to words, natural-language and cipher models inherit one set of constraints. If spaces are introduced by the encoding system, different rules apply. If token generation is template-based, spacing may simply mark the completion of one generated form. If the system copies and mutates nearby tokens, spaces become boundaries of reusable units.

A mature mechanism should therefore explain not merely token contents but why token boundaries occur where they do.

6. Voynich words have shape families

One of the manuscript’s strongest mechanical impressions is that many tokens look like near neighbours. A form appears, then another differs by a small substitution, extension, deletion or prefix-like change.

This behaviour is compatible with morphology in natural language. It is also compatible with abbreviation families, slot grammars and copy-and-modify generation. The observation therefore does not solve the manuscript, but it creates a central gate: any successful mechanism must explain why the token space is densely populated by related forms.

Our earlier Word Families work already established this as a research spine. Here it becomes machinery.

7. The token is not assembled freely

If Voynich glyphs were combined with complete freedom, the number of possible tokens would explode. The actual manuscript occupies a much smaller structured portion of that possibility space.

Certain glyphs prefer early positions. Others prefer late positions. Some combinations recur. Others are scarce. This resembles a grammar of token construction even before we decide whether that grammar is linguistic.

A slot-based mechanism naturally predicts positional restriction. Natural morphology can do the same. Abbreviation practices can create stereotyped endings. Cipher transformations may preserve some source-language structure depending on the cipher. A free random generator would need extra constraints added deliberately.

8. Prefix-like behaviour

Some Voynich forms act as though the beginning of a token is drawn from a restricted set of options. In EVA transcription, strings beginning with combinations such as q- have often been treated as a distinctive structural class.

Calling these prefixes would imply linguistic function. MECHANISM instead asks whether they are initial operators: recurrent beginnings that alter the probability of what follows.

An initial operator could represent morphology, abbreviation, cipher state, token template selection or scribal production habit. The important measurement is conditional dependence: given the beginning, how constrained is the rest?

9. Suffix-like behaviour

Token endings are also constrained. Certain terminal forms recur disproportionately. That can look like inflectional morphology, but the mechanism question is broader.

Terminal choices might close a token template. They might encode grammatical endings. They might result from the shapes that are easiest to append during local modification. They might respond to line space or following context.

A powerful test therefore asks whether endings depend only on the token itself or also on what comes before and after. Cross-token dependence would distinguish some mechanisms from memoryless token builders.

10. The middle of the word can be a machine too

It is tempting to model Voynich tokens as a prefix, a core and a suffix. That is useful, but we should not assume the core is semantically lexical.

A middle component may function as a stem, template choice, copied shape, state marker or compressed code group. If particular central forms accept only particular beginnings or endings, then the token behaves like a constrained product of interacting components.

The key quantity is not visual neatness. It is predictive power. Does the proposed decomposition predict held-out tokens better than a simpler model?

11. Token length is part of the mechanism

Voynich tokens occupy a relatively restricted length range. Extremely long strings are uncommon. Very short units exist but are constrained.

Every mechanism family must account for this distribution. Natural languages have characteristic word-length profiles. Abbreviation can shorten them. Cipher systems can preserve or expand length. Template generators often impose hard or soft bounds. Copy-and-modify can preserve a stable length distribution automatically if edits are local.

Length therefore acts as a quiet but useful discriminator. A mechanism that generates the right-looking glyphs but produces too many long tokens has failed.

12. Frequency is not free

Some tokens occur often. Many occur rarely. A substantial tail of low-frequency forms exists. This resembles the skewed distributions found in natural language, but similar shapes can arise from other productive systems.

A copy-and-modify mechanism can create a few stable high-frequency ancestors and many rare descendants. A template system can combine common components into many low-frequency products. Natural language generates rare words through vocabulary and morphology.

So frequency shape is not a fingerprint by itself. It is one dimension in a multi-gate test.

13. Zipf-like behaviour is not a verdict

Voynich discussions often invoke Zipf’s law because word frequencies in human language roughly follow a rank-frequency relationship. Voynichese exhibits broad frequency regularities that invite such comparison.

But many generative processes produce heavy-tailed distributions. Preferential reuse, copying, template combination and random processes with constraints can all mimic Zipf-like behaviour.

The mechanism lesson is simple: a statistic that many unrelated engines can reproduce has low discriminating power. Rainbolt would not stop at “this country has mountains.” We should not stop at “this text is Zipf-like.”

14. Character entropy measures constraint

Entropy asks, roughly, how uncertain the next symbol is. Lower conditional entropy means preceding context makes the next choice more predictable.

Voynichese has long attracted attention for relatively strong character-level predictability under some comparisons. This is a real mechanistic clue because it tells us the glyph stream does not wander freely.

But the cause is not determined by the measurement. Strong orthography, token templates, abbreviation, local copying or encoding can all reduce entropy. The right question is which mechanism reproduces the entropy while also passing other gates.

15. Entropy can move when representation moves

Entropy is only as stable as the representation feeding it. Merge two glyph classes and uncertainty changes. Split one glyph into two and it changes again. Alter tokenisation or exclude rare forms and results can move.

This does not make entropy useless. It makes robustness testing essential.

The strongest mechanism claims survive several reasonable transcription choices. If an apparent anomaly vanishes when one ambiguous glyph class is handled differently, it should not carry the weight of a grand theory.

16. Word entropy tells a different story

Character-level predictability can coexist with substantial diversity at the token level. This is important because a system can be highly constrained in how it builds words while still producing many distinct words.

That architecture naturally fits productive morphology, template combination and copy-and-modify mechanisms. It is less consistent with a simple tiny codebook repeatedly emitting the same few forms.

The mechanism must therefore explain both restriction and productivity: why the pieces combine in limited ways while the manuscript still generates a large token inventory.

17. Local repetition is a major clue

Voynich contains striking local repetition. Tokens can repeat exactly or appear as close variants near one another.

Natural language repeats words locally when topics persist. Formulaic technical writing can repeat instructions. Copy-and-modify generation predicts local resemblance directly. A memoryless template generator predicts the global distribution but may underproduce local clustering unless local state is added.

This is why locality is more discriminating than frequency alone. It tells us the next token may depend on recent history.

18. Exact repetition and near repetition are different

A system that repeats exact tokens and a system that repeats token families can have very different machinery.

Exact repetition suggests reuse of a stable unit. Near repetition suggests transformation. The ratio between them can therefore constrain how much copying versus modification occurs.

A copy-and-modify engine naturally creates both: sometimes reproduce the source token, sometimes make one small edit. A morphological language creates related forms according to grammar. A slot system creates near neighbours whenever a component changes.

The next test is whether the edits themselves are structured.

19. Edit distance becomes a microscope

Edit distance counts the minimum insertions, deletions or substitutions needed to turn one token into another. For Voynich, it gives a practical way to quantify family resemblance.

If nearby tokens are unusually close in edit distance compared with randomly paired tokens from the same page or section, local-copy mechanisms gain support. If closeness is fully explained by high-frequency templates, then explicit copying may not be necessary.

The denominator matters. A text with a restricted token grammar will naturally contain many near neighbours globally. The meaningful test is whether proximity adds extra resemblance beyond that baseline.

20. The line is a functional candidate

Prescott Currier drew attention to line-level behaviour decades ago, and later work has continued to treat line position as an important anomaly. Certain forms show strong preferences near line starts or ends.

If a line were merely a modern-looking wrap caused by page width, such effects would be less expected. A functional line suggests the production process knows where the line begins and ends.

Natural writing can still produce line effects through abbreviation, justification or scribal fitting. Copy-and-modify can use the current line as its memory window. Template systems can reset state at line breaks. Cipher procedures can operate line by line. The observation is powerful because every engine must explain it.

21. Line beginnings may reset state

A recurring pattern at the beginning of lines can be interpreted mechanically as a reset, header, initial condition or preferred start form.

This does not mean the line is a sentence. It means the production process may enter a different state when beginning a new line.

A stateful model can test this directly: estimate token or glyph probabilities at ordinary internal positions, then compare them with the first positions after a line break. If the distribution shifts reproducibly, the line boundary carries information.

22. Line endings may be constrained by geometry

At line ends, a scribe faces a physical boundary. This creates mundane mechanisms that can imitate linguistic effects.

A scribe may choose shorter variants, abbreviate, stretch characters, select a different word form or end a unit early. If Voynich endings shift near the margin, part of the mechanism may simply be page geometry.

This is why layout should be included in the model. A purely symbolic analysis can mistake the physical writing surface for grammar.

23. Cross-line behaviour can falsify models

If dependence is strong within a line but weak across a line boundary, the line is not merely visual. The production mechanism is respecting that boundary.

This creates a valuable adversarial test. A model trained only on token frequencies may reproduce word shapes while failing cross-line transition behaviour. A local-copy model with a reset at each line may succeed. A natural-language model might require an explanation in terms of line-level units or scribal formatting.

One of the rules of the new architecture is emerging: do not accept a model because it matches the centre if it fails the boundaries.

24. Paragraphs create another state level

Many pages contain paragraph-like blocks. Some glyph forms and token distributions have been associated with paragraph starts or local paragraph structure.

A paragraph may represent a semantic unit, recipe entry, source segment, copying batch or simply a visual production block. MECHANISM does not decide which. It asks whether probability distributions change when a paragraph begins.

This adds another nested state: glyph inside token, token inside line, line inside paragraph, paragraph inside page.

25. Pages are not independent bags of words

Voynich pages differ measurably in vocabulary and character-pair behaviour. Page identity therefore matters.

This can arise from subject matter, scribe, production phase, local source or a page-level state variable in the generative process. The important point is that the manuscript’s statistics cannot always be understood by pooling the entire corpus.

Pooling can manufacture apparent complexity. A mixture of two simple page types can look like one complicated global system. Conversely, analysing pages separately can reveal stable rules hidden by aggregation.

26. Currier A and B are mechanism states

Currier A and Currier B are best treated here as observable corpus states, not as translated languages.

The distinction has survived decades because different folios show consistent statistical differences. A 2026 quantitative study by Christophe Parisel further confirmed that the A/B distinction can be recovered from text statistics without simply copying Currier’s labels, while also arguing that A/B may be a coarse projection of a richer generative structure.

That finding fits the mechanism-first approach perfectly. A/B may not be the machine. It may be one visible setting of the machine.

27. A page-level switch is a powerful idea

If some textual choices are set once per folio and then persist across the page, the system contains a page-level parameter.

That is different from ordinary local word formation. It means a writer or source may enter the page with a particular regime already selected.

A page-level state could correspond to dialect, scribe, source, encoding table, subject category or production convention. MECHANISM does not name the hidden cause. It records the architecture: global settings can constrain local output.

28. The five-scribe model changes the machine

Lisa Fagin Davis’s palaeographic work identifying five hands changes how we should imagine production. The manuscript is not best modelled by default as one operator with one fixed behaviour.

Multiple scribes create a hierarchy: shared system, individual operator settings. The central question becomes which properties are stable across hands and which change with the hand.

If several scribes share the same deep token grammar but differ in surface preferences, that suggests transmission of a common method. If each hand has independent mechanics, the codex may be a compilation. The scribe therefore becomes a natural state variable in every statistical model.

29. Scribe one and Currier A are not interchangeable concepts

Modern palaeographic analysis has revealed relationships between hands and Currier categories, but the concepts remain different.

A hand is graphical production. Currier state is textual distribution. Their correlation is evidence about mechanism because it asks whether operator identity changes the text system.

Where the two align, scribal practice becomes a plausible cause. Where textual state changes without hand change, other hidden variables are required. This is the kind of cross-domain handoff that CivDJ is designed to preserve.

30. Sections introduce content or production state

Herbal, astronomical, balneological, cosmological, pharmaceutical and recipe-like sections are visually different. Their text distributions also show differences.

In a natural-language model, subject matter should alter vocabulary. In a production model, different sections may have been written at different times, from different sources or by different scribes. In a mixed mechanism, both can occur.

The correct comparison therefore controls for section before interpreting global statistics. A signal that disappears within sections may have been generated by corpus mixture rather than local semantics.

31. Burstiness can be real and still misleading

Rare words in meaningful texts often cluster by topic. This clustering, sometimes called burstiness, has been used as evidence that Voynichese may contain semantic organisation.

But section drift can create the same global effect. If one vocabulary is used heavily in one section and another elsewhere, rare tokens appear bursty even without sentence-level semantic dynamics.

This teaches a general mechanism lesson: a macro-pattern can be an artefact of mixing micro-regimes. Analyse at multiple scales before assigning cause.

32. Adjacent-word dependence is a critical gate

Meaningful language usually creates dependencies between neighbouring words. Grammar, syntax and semantics constrain what can follow what.

If Voynich adjacent-token dependence is weak after controlling for token frequencies, simple slot-generation mechanisms gain plausibility. If strong cross-token dependencies survive controls, memoryless token builders fail.

This is one of the most valuable tests because it asks whether the machine finishes one token and forgets it, or carries state forward.

33. Memoryless generators are easy to build

A memoryless token generator can learn the distribution of prefixes, cores and suffixes and emit convincing-looking Voynich words. This is useful because it establishes a baseline.

If a simple slot grammar reproduces character entropy, word lengths and common shapes, those features cease to be strong evidence for semantics by themselves.

The question then moves outward. Does the generator reproduce local repetition, line boundaries, page states, scribal differences and cross-word dependencies? Baselines become valuable precisely because they remove easy victories.

34. Template generation explains some things naturally

A template model says tokens are assembled from a small set of allowed components or slots. This immediately explains strong positional restrictions and families of similar words.

It can also generate a large vocabulary from relatively few parts and reduce character entropy without requiring semantics.

Its weakness appears when output depends on recent local context in ways not encoded by the template. A purely independent slot model has no memory. If real Voynich shows excess local similarity or cross-token state, the template must be extended.

35. Copy-and-modify explains locality naturally

A copy-and-modify mechanism starts from an existing token and creates a nearby form by changing one or a few components.

This naturally produces families, local similarity and a long tail of rare variants. If the source token is chosen from a recent window, it also produces local repetition without requiring a global dictionary.

The challenge is explaining why edits obey the manuscript’s deeper positional grammar. Random mutation can damage token structure. A successful copy model therefore needs structure-preserving edits or an acceptance rule that rejects illegal forms.

36. Free-character mutation is too weak a mechanism

If a scribe simply copies a nearby word and changes any character at random, the output tends to drift into combinations that do not resemble the manuscript’s preferred forms.

This matters because it tells us that “copying” is not a complete explanation. The mutation process itself must be constrained.

Perhaps only visually similar glyphs substitute. Perhaps edits occur in slots. Perhaps the writer remembers legal word-shapes. Perhaps the source material imposes morphology. Whatever the cause, the system preserves structure while varying tokens.

37. Structured local copy is more interesting

A stronger copy model combines two ideas: reuse recent forms and restrict changes to legal transformations.

This creates a plausible mechanism for the manuscript’s strange combination of productivity and sameness. The writer can generate many distinct tokens while remaining inside a narrow structural manifold.

Recent 2026 exploratory computational work has pushed this class of model further, arguing that structure-preserving local copying can reproduce several observed signatures simultaneously. Such work is promising as a mechanism test, but it remains a research proposal rather than a settled explanation. Its real value is the discriminating predictions it creates.

38. The natural-language engine remains alive

Nothing in mechanism-first analysis requires us to abandon natural language.

Languages have morphology, spelling constraints, word families, page-level vocabulary changes and local dependencies. Medieval technical language can be highly formulaic. Abbreviations can further compress and regularise the surface.

The natural-language model becomes weak only when it is invoked vaguely. It becomes strong when a specific language and writing convention predict measurable Voynich behaviour that alternatives do not.

39. Abbreviation is a bridge mechanism

Medieval abbreviation practices are important because they can make ordinary language look mechanically unusual.

Repeated common endings may be compressed. Frequent words may receive compact forms. Context can determine expansion. Scribal systems can develop specialised shorthand within professions or institutions.

An abbreviation-heavy manuscript could therefore exhibit strong token constraints while still encoding natural language. This mechanism sits between direct-language and cipher families, reminding us that our six families are a search scaffold rather than six sealed boxes.

40. Cipher models must explain surface structure

A cipher is not validated merely because it yields readable text from selected passages.

The transformation must also explain why the visible ciphertext has the observed token grammar, line effects, repeated families and page variation.

Simple substitution preserves many source-language frequency relationships. More elaborate ciphers can destroy or transform them, but every extra mechanism requires historical plausibility and stable rules. If the cipher changes method whenever a difficult passage appears, it has become an explanation generator rather than a decipherment.

41. Polyalphabetic and stateful ciphers remain possible

A stateful cipher can change mappings over time or across pages. This could in principle produce Currier-like variation or page-level regimes.

But the historical and operational cost rises. The more complicated the cipher, the more procedure a fifteenth-century writer must execute reliably across hundreds of pages and multiple scribes.

This operational burden is itself evidence. A mechanism must not only fit the statistics; it must be performable by the humans and tools available in the production environment.

42. Human executability is a mechanism gate

The manuscript was physically written by human hands. Any production mechanism therefore has an operator cost.

Could a scribe execute the rule repeatedly without a computer? Does it require large lookup tables? Does it require remembering long histories? Can several scribes learn the same procedure? Would the procedure be faster or slower than ordinary writing?

A mechanism that reproduces the corpus computationally but would be nearly impossible for a historical scribe to execute should lose confidence unless evidence supports the necessary apparatus.

43. Production speed leaves traces

Writing hundreds of pages is labour. Mechanisms differ in production cost.

Direct copying is relatively fast. Enciphering every symbol through a complex table is slower. Generating pseudo-text from local visual rules may be fast once learned. Abbreviation can increase speed.

Stroke fluency, correction frequency and layout can therefore become indirect evidence about how cognitively expensive the system was to execute. A future MECHANISM Master should connect palaeographic observations with computational procedure instead of treating them as separate worlds.

44. Multiple scribes imply transmissibility

If five scribes used the same deep writing system, the mechanism had to be teachable or copyable.

This is a strong constraint. A private idiosyncratic code invented in one person’s head becomes less plausible unless the other hands were simply copying existing text mechanically. A shared production method, by contrast, can be transmitted through instruction, exemplars or institutional practice.

The sociological mechanism matters. Machines do not exist outside operators.

45. Copying from an exemplar is different from generating text

A scribe can produce structured Voynich text in two fundamentally different ways: generate it or copy it.

If copying from an exemplar, many observed quirks could originate in an earlier source. Scribal hands would then tell us about transmission rather than composition. If generating directly, each scribe must know the production rules.

This distinction can sometimes be approached through corrections, repeated errors, line fitting and relationships among hands. A copied text and a generated text leave different operational signatures.

46. Text-image order is another hidden mechanism

Did the maker draw first and write around the image? Write first and add imagery later? Plan both together? Use different operators?

The answer matters because image placement can constrain text shape and line length. If text avoids drawings, some line effects may be geometric rather than linguistic. If labels are added after images, they may belong to a different micro-process from running text.

The manuscript therefore contains several overlapping machines: writing, illustration, colouring, page layout and binding.

47. Labels may be a separate sub-machine

Short strings beside stars, diagram elements and plant structures often behave visually like labels.

If labels are generated differently from running text, pooling them can confuse statistics. They may use names, abbreviations or special grammatical forms. They may also be mechanically copied from source diagrams.

A serious model should test labels as their own corpus and then ask how their structure connects to paragraph text.

48. The Rosettes foldout may demand another operating mode

Large diagrammatic foldouts create a physical environment unlike ordinary pages. Text appears around spatial structures rather than in simple paragraph flow.

This can alter production mechanics. Labels may be placed after visual planning. Reading order may be radial or relational. Space may encode adjacency.

A single universal left-to-right sentence model may therefore be inappropriate for every part of the codex. MECHANISM must be willing to admit multiple operating modes while still searching for shared deep rules.

49. The mixed-system hypothesis becomes increasingly natural

As more layers accumulate, a mixed system begins to look less like an escape hatch and more like a serious structural possibility.

Natural language could be abbreviated. Abbreviated text could be enciphered. Labels could use a different convention from paragraphs. Some sections could be copied while others were compiled. Scribes could share a deep system while varying surface habits.

The danger is unlimited complexity. A mixed model can explain anything if every anomaly receives its own mechanism. The antidote is compression: prefer the smallest number of rules that explains the largest number of independent observations.

50. Mechanism quality is compression quality

A strong mechanism earns trust by replacing many observations with a smaller set of rules.

If one rule explains token families, low character entropy, local similarity and line behaviour, it is more valuable than four unrelated stories that each explain one feature.

This is close to minimum-description-length reasoning: the best model balances fit against complexity. A model that memorises the manuscript perfectly has learned nothing if its rulebook is as long as the manuscript.

51. The first mechanism tournament

We can now stage a first tournament across the six families.

• Can it reproduce glyph-position constraints?
• Can it reproduce token families?
• Can it reproduce the word-length distribution?
• Can it reproduce frequency shape?
• Can it reproduce character entropy?
• Can it reproduce local repetition?
• Can it reproduce line-start and line-end effects?
• Can it reproduce paragraph behaviour?
• Can it reproduce page-level states?
• Can it reproduce Currier A/B differences?
• Can it survive multiple scribes?
• Can it operate across sections?
• Can a fifteenth-century human execute it?

No mechanism should be declared victorious because it wins one column.

52. Natural language: strengths and weak points

Natural language naturally explains meaningful vocabulary, morphology, local dependence and section-specific words. It also fits the historical expectation that a manuscript usually contains information intended for humans.

Its weak points are the manuscript’s unusual surface regularities and the lack of a stable language identification. A successful language model must explain those anomalies through a specific orthography, abbreviation system or transformation rather than hand-wave them away.

53. Abbreviation: strengths and weak points

Abbreviation naturally fits medieval scribal culture and can generate compressed, repetitive-looking forms. It can also create strong positional structure if conventional marks cluster at word endings.

Its challenge is scale. The system must explain the entire corpus consistently, not merely interpret a few glyphs as familiar shorthand. It also needs a plausible connection between the abbreviation repertoire and the manuscript’s unusual token families.

54. Cipher: strengths and weak points

Cipher explains intentional unreadability directly. It fits the modern intuition that an unknown script may hide known language.

Its weakness is over-flexibility. Many proposed ciphers can extract selected readable words while failing global structure. A real cipher solution should reduce ambiguity as rules accumulate, not require more exceptions.

55. Template generation: strengths and weak points

Template generation is excellent at token grammar, character predictability and productive families. It gives a compact explanation for why words look related.

Its weakness is long-range and local context. A memoryless template may produce plausible individual words while failing the order in which real words appear. To survive, it may need page state, line state or local memory.

56. Copy-and-modify: strengths and weak points

Copy-and-modify naturally explains near-neighbour forms and local clustering. It is human-executable and can generate a large vocabulary cheaply.

Its weakness is structural drift. Without a constrained edit grammar, it produces illegal-looking forms. It also needs to explain page-level regimes and any dependencies not reducible to recent visual copying.

57. Mixed system: strengths and weak points

A mixed system can explain heterogeneity without forcing every feature through one mechanism.

Its weakness is exactly that flexibility. If every contradiction triggers another sub-system, the theory becomes unfalsifiable. Mixed models need explicit boundaries: which mechanism operates where, how transitions occur and what predictions distinguish the mixture from one simpler engine.

58. The best current state is not a winner but a narrowing

Mechanism-first analysis does not yet justify declaring one family the complete answer.

What it does justify is rejecting weak versions. Pure random character generation is too unconstrained. A memoryless model that ignores line and page effects is incomplete. A cipher that depends on ad hoc exceptions is weak. A language identification that cannot reproduce surface structure is weak. A copy process that mutates freely is weak.

The remaining space is smaller and more structured than the phrase “anything is possible” suggests.

59. What Rainbolt adds to mechanism research

Rainbolt’s most transferable lesson is not image recognition. It is clue discrimination.

For Voynich, the question is not which feature is most famous. It is which feature best separates candidate engines.

Zipf-like frequency may be weak because many mechanisms produce it. A line-boundary dependency may be strong because fewer mechanisms reproduce it. A page-level switch may be stronger still if it forces an otherwise simple generator to become stateful.

Mechanism research should therefore rank tests by information gain.

60. The next discriminating experiment

At any moment, the best experiment is the one most likely to reorder the mechanism leaderboard.

If language and copy models predict different cross-line dependence, test that. If template and copy models predict different local edit-distance clustering after conditioning on token frequency, test that. If scribal identity and subject matter predict different page-state changes, compare their boundaries.

The Receiver Gauge logic applies directly: do not request more generic information when one targeted observation can separate states.

61. Mechanism research needs matched synthetic controls

One of the most powerful modern tools is the synthetic control: generate text from a proposed mechanism and compare it against the manuscript using the same metrics.

This turns a verbal theory into an executable claim.

If a slot grammar is proposed, generate a corpus of equal size. If local copying is proposed, implement the copying rule. If a cipher is proposed, encipher matched historical text. Then compare not one statistic but a panel of signatures.

Mechanisms that only sound plausible become measurable.

62. One statistic is never enough

A generator can be tuned to hit one target almost trivially.

Match entropy and it may fail word diversity. Match word diversity and it may fail local repetition. Match local repetition and it may fail line boundaries. Match Currier states and it may fail scribal hand relationships.

A mechanism deserves attention when it lands inside several independent constraints without being separately tuned to each one.

63. Hold-out testing is the decipherment equivalent of an examination

A theory built on every page can overfit every page. The cure is hold-out data.

Develop the mechanism on one set of folios. Freeze the rules. Then predict unseen folios.

This is the same reason students sit examinations with unseen questions. Competence is demonstrated when a rule transfers beyond the examples used to learn it.

A Voynich solution that cannot survive unseen pages has memorised the mystery rather than explained it.

64. Cross-scribe testing is even harder

Train on one scribe and test on another.

If the deep mechanism transfers while surface frequencies shift, we may have found a shared production grammar. If the model collapses, scribal identity is doing more work than expected.

This is one of the most important future directions because multiple hands let the manuscript provide its own internal replication environment.

65. Cross-section testing separates topic from machinery

A model that only works in herbal pages may be modelling botanical vocabulary or one production batch rather than the manuscript’s general mechanism.

Train on one visual section and test another. Stable low-level rules suggest shared machinery. Section-specific failures reveal where subject, source or operator state enters.

This converts the visual divisions of the codex into natural experiments.

66. Cross-transcription testing protects against analyst artefacts

Run the same analysis across reasonable transcription variants.

If a mechanism only appears under one symbol convention, it may be modelling the transcriber. If the signature survives several encodings, it is more likely to belong to the manuscript.

This is a simple but essential provenance gate for eduKateAI.

67. The manuscript is a hierarchy of states

We can now summarise the machine as a hierarchy.

• stroke state;
• glyph state;
• token state;
• neighbourhood state;
• line state;
• paragraph state;
• page state;
• Currier state;
• scribe state;
• section state;
• codex state.

The central research task is to identify which level controls which observable choice.

68. Hidden state is the bridge between appearance and cause

When two pages look statistically different, some hidden state changed.

The hidden state could be topic, scribe, language, source, cipher table, template regime or production phase. Mechanism research does not need to know the name immediately. It can first infer how many states are required and where transitions occur.

This is precisely how modern sequence models work: observable outputs are used to infer latent structure. Voynich invites the same discipline.

69. Boundaries carry disproportionate information

Lines, paragraphs, pages, quires, hands and sections are all boundaries.

At each boundary we can ask whether the distribution resets, drifts or continues. The answer tells us the memory length of the process.

A mechanism that carries state across words but resets at lines is different from one that carries state across pages. Boundary analysis therefore compresses many research questions into one general method.

70. The centre-and-edge handoff appears inside the manuscript

Every discipline has a centre where it owns the evidence and an edge where it must hand off.

Statistics can identify a page-level switch but cannot name its historical meaning. Palaeography can identify hands but not automatically explain token entropy. Linguistics can test morphology but cannot date parchment. Codicology can reconstruct gatherings but cannot translate labels.

MECHANISM works only when those partial owners exchange state without pretending to own the whole answer.

71. The machine should generate predictions, not just explanations

An explanation looks backward: it tells us why observed patterns make sense.

A mechanism should also look forward.

If local copying is real, nearby unseen tokens should show predictable similarity. If a page-level switch governs certain substitutions, the switch should classify held-out pages. If a scribe-specific grammar exists, new hand assignments should predict textual tendencies.

Prediction converts plausibility into risk. A theory that risks being wrong can earn confidence.

72. Good mechanisms become narrower as they succeed

At the beginning, a hypothesis may have many free choices. As evidence accumulates, those choices should freeze.

A cipher mapping should stabilise. A token grammar should stop adding new slots. A copy rule should stop changing window size. A language hypothesis should use the same morphology across pages.

If the mechanism requires increasing flexibility to absorb new evidence, apparent success is hiding failure.

73. Failure signatures are as valuable as successes

Suppose a slot grammar matches character entropy but fails local word order. That failure localises the missing mechanism.

Suppose a copy model matches local similarity but cannot reproduce Currier A/B. That tells us an additional global state is required.

Failure therefore becomes architectural information. It tells us which layer the model lacks.

74. A mechanism can be partly right

The final Voynich explanation may contain pieces of several theories that are currently presented as competitors.

A slot grammar may correctly describe surface token construction even if the tokens encode language. Local copying may correctly describe scribal production even if the source text is meaningful. Abbreviation may explain endings while a cipher explains glyph values.

This is why failed global theories can preserve successful local mechanisms.

75. The Ouroboros danger in mechanism research

A dangerous research loop occurs when a hypothesis generates a synthetic corpus, the synthetic corpus is used to define what counts as Voynich-like, and then similarity to that definition is treated as evidence for the hypothesis.

The model has begun eating its own assumptions.

Controls must therefore be independent where possible. Metrics should be selected before looking at model output. Hold-out tests should remain untouched. Historical plausibility should come from external evidence rather than the mechanism itself.

76. AI makes mechanism search dangerously cheap

Modern AI can generate thousands of candidate grammars, transformations and historical narratives far faster than humans can evaluate them.

This changes the bottleneck. Idea generation is no longer scarce. Falsification is scarce.

eduKateAI should therefore spend disproportionate computation on controls, held-out prediction, provenance and failure analysis rather than merely generating more candidate decipherments.

77. The MECHANISM Master in eduKateAI

This article should become the public reasoning surface for a Voynich MECHANISM Master.

Internally, the Master should maintain observable signatures and mechanism families separately. Signatures include token grammar, entropy, local similarity, line effects, page states, Currier variation, hand variation and section variation. Mechanism families propose procedures capable of generating those signatures.

The Master’s job is not to announce a solution. It ranks models by the number of independent gates passed, penalises complexity, tracks failed signatures and sends unresolved causes to the correct domain Masters.

78. Typed edges for the machine

The underlying knowledge graph should use relations such as:

• PRODUCES — mechanism produces observed signature;
• TRANSFORMS — one state changes into another;
• CONSTRAINS — evidence reduces allowed mechanism space;
• VARIES_WITH — feature changes with hand, page or section;
• DEPENDS_ON — output uses prior state;
• RESETS_AT — state boundary;
• FAILS_TO_REPRODUCE — model misses a signature;
• COMPATIBLE_WITH — observation fits but does not discriminate;
• HANDOFF_TO — unresolved cause belongs to another Master;
• RETURNS_TO — tested claim goes back to KNOWLEDGE.

These edges let AI traverse mechanism rather than merely retrieve prose.

79. The mechanism leaderboard

A future interface could display mechanism families against a live set of gates.

Green means reproduced under matched conditions. Amber means compatible but not discriminating. Red means failed. Grey means untested.

The important design principle is that no one overall confidence number should erase the shape of the evidence. A model may be strong on token formation and weak on page states. Another may explain lines but fail historical executability.

80. What would count as a breakthrough mechanism?

A breakthrough would not merely make Voynich-like text.

It would reproduce several difficult signatures with few rules, survive hold-out folios, transfer across scribes or explain why it should not, operate across sections or specify boundaries, remain robust to transcription choices, be historically executable and generate new predictions that the real manuscript subsequently satisfies.

Only then would the mechanism begin to deserve priority over alternative engines.

81. A breakthrough mechanism still would not equal translation

Even if we discovered exactly how the surface text was generated, semantics could remain unresolved.

Imagine proving that tokens are produced through a constrained abbreviation grammar. We would still need to know what the underlying abbreviations expand to. Imagine proving local copy-and-modify. We would still need to know whether the copied forms encode words, categories or nothing semantic.

This distinction protects the sequence. MECHANISM owns process. KNOWLEDGE owns the factual floor. Later articles can ask what content and purpose might sit behind the process.

82. The return path to KNOWLEDGE

Every mechanism experiment should end with a return question:

Did this experiment change what we can responsibly say, or did it merely create a better hypothesis?

If it changed the factual state, update KNOWLEDGE with provenance. If it only improved a candidate mechanism, keep it attached to the mechanism branch. This prevents exploratory models from silently becoming “facts about Voynich” after enough repetition.

83. What the manuscript is doing, in one sentence

The safest mechanism summary today is this:

The Voynich Manuscript produces highly constrained word-like forms through rules that operate at several nested scales, with measurable variation across local context, lines, pages, textual states, scribes and sections.

That sentence is not a decipherment.

It is already enough to reject many weak models.

84. The deeper machine may be simpler than the surface mystery

Mysteries feel complex because we do not know which variables matter.

Once the correct state variables are found, an apparently impossible surface can collapse into a compact mechanism. Currier A/B may reduce to a small set of parameter shifts. Token families may reduce to a handful of legal transformations. Line effects may reduce to reset rules.

The purpose of MECHANISM is to discover those reductions without prematurely assigning meaning.

85. Why Voynich is an ideal machine-learning benchmark

The manuscript offers a rare benchmark where the surface data is rich but ground truth is incomplete.

An AI system cannot simply optimise toward a known translation. It must demonstrate disciplined behaviour: model comparison, uncertainty, robustness, provenance, held-out prediction and failure tracking.

That makes Voynich valuable beyond medieval studies. It tests whether AI can reason when the world does not supply a neat answer key.

86. Why this transfers directly to education

A student’s wrong answer is also an output from a hidden mechanism.

The surface answer may be 42. The hidden cause could be misunderstood ratio, sign error, copied number, wrong formula, language confusion or carelessness. A weak tutor corrects the output. A strong tutor infers the mechanism and asks the next discriminating question.

Voynich and tuition therefore share the same architecture: output → candidate mechanisms → discriminating test → repair or update.

87. Why this transfers to news and finance

A market price is an output from many interacting mechanisms. A news narrative is an output from information acquisition, selection, framing and distribution.

In both cases, looking only at the surface can mislead. Mechanism-first reasoning asks which process could have generated the observable pattern and what alternative mechanisms predict differently.

This is why the Voynich machine belongs inside the wider eduKate civilisation map.

88. Why this transfers to institutions

An institution is also a generative system. Rules, incentives, handoffs, people and memory produce recurring outputs.

If an institution repeatedly fails at the same boundary, the surface incidents may be different while the mechanism is stable. Good analysis searches for the recurring transform rather than blaming each output independently.

Voynich trains the same habit at a smaller, stranger scale.

89. The machine we cannot yet read is still readable as a machine

This is the central paradox of the article.

We cannot securely read the manuscript semantically.

But we can read its constraints.

We can see that certain forms belong in certain positions. We can see local families. We can measure boundary effects. We can distinguish page states. We can relate scribal hands to textual variation. We can build synthetic competitors and watch them fail.

Mechanism is a kind of reading that comes before translation.

90. The return

At the beginning we asked what kind of machine could produce the observable Voynich text.

We are not yet entitled to one final answer.

We are entitled to a better specification.

• The machine preserves strong glyph and token constraints.
• It produces dense families of related word-like forms.
• It combines restriction with productivity.
• Recent context matters.
• Line boundaries matter.
• Page-level state matters.
• Currier A/B captures real but incomplete variation.
• Scribal hands must be part of the model.
• Visual sections and production layers matter.
• Simple memoryless models are insufficient as complete explanations.
• Human executability and historical plausibility are required.
• Any winning mechanism must survive several independent signatures at once.

This is no longer an unstructured mystery.

It is a machine with a growing test specification.

---

Primary routes from MECHANISM

• Voynich Research Library — the full experimental Warehouse.
• Voynich Longform 01 · HUMAN — the receiver and premature certainty.
• Voynich Longform 02 · KNOWLEDGE — the factual floor.
• Word Families — token similarity and productive structure.
• The Control Problem — matched comparisons and false anomalies.
• Voynich | The Mathematics and Science — the quantitative owner.
• What a Real Voynich Decipherment Must Survive — adversarial gates.

External research anchors

Mechanism research benefits from returning to primary and specialist sources rather than repeating popular summaries. Particularly useful anchors include Prescott Currier’s original statistical observations, René Zandbergen’s maintained text-analysis reference, modern palaeographic work on scribal hands, and recent quantitative attempts to recover Currier structure computationally.

• Prescott Currier · Some Important New Statistical Findings
• René Zandbergen · Voynich MS Text Analysis
• Christophe Parisel · A Quantitative Confirmation of the Currier Language Distinction (2026)

Next: Article 4 changes the geometry. MECHANISM asks how one observable system might operate. MIRROR / OUROBOROS will place opposing explanations on the same evidence and ask how the same clues can produce mutually incompatible stories — and how a research system prevents itself from feeding those stories back as facts.