Voynich Longform 07 · CASE STUDY

Voynich Manuscript | Case Study: The Self-Citation Hypothesis

One of the strongest ways to test a research architecture is to stop speaking in abstractions and force one real hypothesis through every gate.

This is that test.

The hypothesis is usually associated with Torsten Timm and Andreas Schinner: the visible Voynich text may have been produced, at least substantially, through a process in which existing word-like forms are copied, modified and reused locally. The mechanism has been described in terms of self-citation, local copying and structured mutation.

The point of this article is not to crown the hypothesis.

The point is to find out exactly how far it travels before it stops.

A useful case study does not ask whether a theory sounds plausible. It asks which parts survive contact with the whole system.

1. Why choose self-citation for the first case?

Because it is concrete.

Unlike a vague “maybe it is encoded language” story, self-citation proposes an observable production procedure. A scribe selects an existing form, copies it, changes it in a constrained way, and thereby creates another token.

2. Concrete mechanisms can fail clearly

That makes the hypothesis scientifically useful even before any verdict.

If a mechanism can be executed, it can be simulated. If it can be simulated, it can be compared with the manuscript under frozen metrics.

3. The original attraction

Voynichese contains dense families of similar words. Exact repetitions and near repetitions appear locally. Small changes can turn one common form into another.

Self-citation makes those patterns cheap to produce.

4. The mechanism in plain language

look at a recently available word-like form;
copy it exactly or approximately;
insert, remove, replace or rearrange a small component;
optionally combine parts of two visible forms;
write the result;
allow the new result to become future source material.

5. Why this creates families automatically

If each new token descends from an older token by a small transformation, the corpus naturally becomes a network of near neighbours.

No dictionary is required to create family structure.

6. Why this creates a long tail

Frequently reused source forms remain common. Occasional mutations create many rare descendants.

This naturally produces a few frequent types and many rare ones.

7. Why this can create local repetition

If the source pool is local, recent forms are disproportionately reused.

That creates clusters of exact and near repetition without requiring topical semantics.

8. The first important result: resemblance is no longer diagnostic

Once a simple local-copy mechanism can produce word families, the mere existence of word families no longer proves morphology.

This is already a major contribution of the hypothesis.

9. The hypothesis attacks an intuition

Humans see structured variation and instinctively infer meaningful word formation.

Self-citation demonstrates that structured variation can also emerge from production history.

10. That does not prove meaninglessness

This distinction matters immediately.

A meaningful text can also be copied and modified. A scribe can work from a meaningful exemplar while preserving local visual relations. Surface copying and semantics are not mutually exclusive.

11. CASE rule one: separate mechanism from interpretation

“Local copying occurs” is a mechanism claim.

“Therefore the manuscript is meaningless” is a semantic conclusion requiring additional evidence.

12. The Timm-Schinner contribution

Timm and Schinner published an executable generating algorithm in Cryptologia, building on earlier work by Timm concerning word similarity and co-occurrence. Their model is important because it moves beyond verbal speculation into a reproducible mechanism family.

The model belongs in the test harness regardless of whether its strongest interpretation survives.

13. What the hypothesis predicts cheaply

many near-neighbour tokens;
local clustering of similar forms;
frequent reuse of common forms;
a long tail of rare descendants;
structural inheritance across neighbouring tokens;
human-executable generation without a large lexicon.

14. What it does not automatically predict

line-start and line-end effects;
page-level state differences;
Currier A/B structure;
cross-scribe invariants;
section-specific regimes;
image-text coupling;
historical purpose;
semantic content or its absence.

15. The case begins with token families

This is the strongest home territory for self-citation.

Near-neighbour structure follows almost directly from the mechanism.

16. Edit distance is the first microscope

If nearby tokens are significantly closer in edit distance than matched random pairs, a local-copy mechanism gains support.

The key control is global family density: Voynich already contains many similar words overall.

17. Local excess matters more than global resemblance

The hypothesis predicts extra similarity because of physical or cognitive proximity, not merely because the lexicon is constrained.

18. The memory window becomes testable

If a scribe copies recent forms, similarity should decay as textual distance grows.

The shape of that decay estimates the effective memory window.

19. A very long memory changes the theory

If source words can be selected from anywhere on the page or manuscript, local clustering becomes less distinctive.

The mechanism gains flexibility but loses predictive sharpness.

20. Source selection is hidden machinery

The theory needs a rule for choosing which word to copy.

Nearest visible word, same line, word above, common form, similar shape, or any recent token are different models.

21. A vague source rule is a failure risk

If the model is allowed to pick whichever source makes the observed output easiest to explain, the case becomes retrospective fitting.

22. The modification rule matters equally

Random free-character mutation is too destructive.

Voynich tokens stay inside a narrow family of legal-looking structures.

23. Structured mutation improves the case

Restrict edits to common substitutions, prefix-like changes, suffix-like changes or known component families, and synthetic tokens remain more Voynich-like.

24. But structure has now entered from somewhere

The moment edits are restricted to legal token transformations, the model requires a grammar of legal transformations.

That grammar is additional machinery.

25. This is the first major pressure point

Self-citation explains variation cheaply.

It does not by itself explain why the variations remain so structurally coherent.

26. A slot grammar can rescue coherence

If tokens contain preferred positions or components, copying can happen inside that grammar.

This hybrid is mechanically stronger than unconstrained mutation.

27. But the theory has changed

We are no longer testing pure self-copying.

We are testing self-copying plus template structure.

28. CASE rule two: name every added mechanism

Hybrids are legitimate.

Hidden hybrids are not.

29. Frequency distribution is supportive but weak

Preferential reuse of common sources can generate skewed frequency distributions and long tails.

This means Zipf-like behaviour is compatible with self-citation.

30. Compatibility is not discrimination

Natural language, templated generation and other systems can also produce heavy-tailed frequency curves.

Frequency therefore does not strongly separate the candidates.

31. Character entropy is another partial success

Structured copying can reduce uncertainty because new forms inherit old structure.

Recent exploratory work has shown that copy-based generators can reproduce low-entropy regimes when mutation respects existing similarity classes.

32. Low entropy still does not identify the cause

Template grammars, abbreviations, ciphers and highly constrained orthographies can also lower entropy.

33. The case strengthens when several signatures align

The real value appears if one frozen copy mechanism simultaneously reproduces token families, low entropy, frequency shape and locality.

34. Multi-signature success is harder to fake

Every added independent signature reduces the chance that the match is accidental or benchmark-specific.

35. Now the case reaches line structure

Voynich line beginnings and endings behave differently from internal positions.

A simple copy mechanism has no reason to care about line boundaries unless line state is explicitly added.

36. Line-reset copying is a plausible extension

The source pool can reset or change at each line.

The scribe might preferentially copy words from the line above or from similar horizontal positions.

37. Vertical copying is a real discriminating idea

Timm has argued that visually similar glyph groups appear above one another often enough to support the idea of copying from nearby lines.

This is stronger than generic local copying because it predicts spatial directionality.

38. Directionality should be tested against geometry

If vertical similarity exceeds horizontal or matched random proximity under controls, the copy mechanism gains specificity.

39. Line endings create another complication

Page width can cause shortening, abbreviation or different token choices.

A copy model may reproduce line effects only after geometry enters.

40. The model is becoming a stack

local copying;
structured mutation;
token grammar;
line-state rules;
page geometry.

41. Complexity is not yet collapse

If each added layer is independently demanded by evidence and improves held-out prediction, the model can legitimately grow.

42. But complexity must pay rent

Every new rule must explain enough new evidence to justify the extra freedom.

43. Page-level states are a tougher test

Different Voynich pages use different token distributions and structural regimes.

A purely local copy process can create drift, but it must explain why some page-level states remain coherent and repeat across sections.

44. Drift is the natural self-citation answer

If each generation inherits local structure, small changes can accumulate over time.

Different regions of the corpus may therefore occupy different points along a structural trajectory.

45. Recent 2026 exploratory work strengthens this part

A recent independent computational project reported a robust structural gradient across the manuscript and found that self-copying models can reproduce important low-entropy gradient behaviour. The same work also found that calibrated local-copy models fail to reproduce the full vocabulary structure at the most extreme pole without additional template machinery.

This is exactly the kind of result a case study needs: partial confirmation plus a precise residual failure.

46. The gradient result is not a final verdict

The 2026 work is exploratory and explicitly calls for external replication.

It should therefore be treated as a strong current constraint, not as settled field consensus.

47. Currier A/B enters the case

Currier A and B are not simply random page-to-page drift. They are recoverable statistical states.

A self-citation model must explain how drift produces stable enough regimes to be detected repeatedly.

48. A regime variable can solve part of this

Different sections or production phases could begin from different source reservoirs or transformation preferences.

Self-citation then operates inside each regime.

49. Again, the hypothesis becomes hybrid

Local copying explains microstructure.

Regime state explains macrostructure.

50. This hybrid is plausible and testable

Train a copy mechanism inside one regime and test transfer into another.

If only a small parameter shift is needed, the shared-chassis idea strengthens.

51. Cross-scribe transfer is crucial

Multiple scribes create a natural question: does the same local-copy grammar survive different hands?

52. If it transfers, the mechanism may be taught or inherited

A shared deep grammar across scribes supports either shared training, shared exemplars or a common production system.

53. If it does not transfer, operator identity dominates

Then self-citation may be individual scribal habit rather than the manuscript’s general mechanism.

54. Current evidence leans toward regime over penmanship

Recent quantitative work suggests structural state differences track section and textual regime much more strongly than scribal identity alone, while individual scribes can traverse different states.

That weakens a simple “each scribe generates their own style” account.

55. It does not weaken shared self-citation

A common local-copy process can still operate under different section-level regimes.

56. The image problem is much harder

Self-citation is a text-generation mechanism.

It does not automatically explain why different kinds of images accompany different textual regimes.

57. Seed-by-page is one possible answer

A page could begin from a label, source word or small reservoir linked to the illustration, then locally drift through copying.

58. This produces a testable prediction

Pages with visually similar content should begin from related token families or occupy similar structural states more often than expected by chance.

59. Labels are therefore the natural bridge

If short diagram labels seed running text, label forms should relate systematically to page vocabularies.

60. But labels are not yet a Rosetta stone

Recent analysis suggests label groups contain section and register structure, but morphology alone does not cleanly show ordinary referent naming.

This leaves image-text coupling unresolved.

61. Meaninglessness is the strongest version of the hypothesis

The most ambitious interpretation says the text can be explained as generated pseudo-text without semantic content.

This is much harder to establish than local copying itself.

62. To prove no meaning, structure must be exhausted

A non-semantic generator would need to reproduce not just local families and entropy, but the full pattern of line, page, section, label and cross-token organisation.

63. Residual structure is therefore decisive

If a calibrated generator still misses coherent token templates, label relations or long-range dependencies, semantic or other structured mechanisms remain live.

64. The strongest current residual is template coherence

Recent exploratory tests report that copy-and-mutation systems can reach the manuscript’s broad low-entropy phenomenology yet produce malformed pseudo-words when pushed toward the most extreme textual regime.

The real text retains coherent word templates that local drift alone does not manufacture cheaply.

65. This does not kill self-citation

It localises what self-citation is missing.

The mechanism may need a slot grammar or other structural acceptance rule.

66. But adding slot grammar weakens the strongest null claim

The more structured the hidden grammar becomes, the less persuasive it is to argue that copying alone explains the manuscript.

67. A grammar can still be meaningless

A generator can enforce legal word shapes without semantics.

Therefore template coherence does not prove language.

68. The mirror alternative remains alive

The same slot structure could represent morphology, abbreviation, encoding or notation.

CASE now hands back to MIRROR.

69. Counterfactual one: ordinary language plus local copying

Imagine that the manuscript contains meaningful language, but the scribe frequently reuses nearby forms because technical vocabulary and morphology are repetitive.

Many self-citation signatures survive.

70. Counterfactual two: abbreviation plus copying

Meaningful source language is compressed into stereotyped components, then locally reused.

This naturally creates low entropy and dense families.

71. Counterfactual three: cipher plus copying

A coded system may preserve or deliberately generate repeated surface forms while underlying content remains meaningful.

72. Counterfactual four: meaningless generator

Local copying plus structural constraints produce pseudo-text designed only to look manuscript-like.

73. All four can share surface copying

Therefore the observation “copying occurred” cannot by itself decide semantic status.

74. CASE rule three: do not let a mechanism inherit a motive

A production mechanism and a historical purpose are separate claims.

75. Historical executability is a strength

Self-citation is human-executable.

A scribe can copy nearby words and alter them without modern computation or elaborate apparatus.

76. Historical motive is weaker

If the text is meaningless, why produce hundreds of pages of carefully structured pseudo-text alongside elaborate illustrations?

Hoax, display, prestige, practice and private systems are possible answers, but each needs evidence.

77. Labour cost becomes evidence

A meaningless-generation theory must account for the substantial material and labour investment of the codex.

78. But labour does not prove semantics

Humans have spent enormous effort on ritual, display, games, art, deception and symbolic systems.

79. Purpose remains open

The case therefore cannot use labour cost as a decisive semantic test.

80. The correction test

If any structurally legal token is acceptable, generation may require few corrections.

If exact semantic or cipher values matter, errors should carry higher cost.

81. Correction scarcity can support low-stakes generation

But fluent meaningful writing by trained scribes can also show few visible corrections.

82. Again the evidence is compatible, not decisive

This repeated pattern is itself informative.

Self-citation often survives as a plausible mechanism while failing to uniquely identify semantics.

83. The held-out examination

A strong case requires frozen rules and unseen material.

Develop the generator on one set of folios, then predict token-family density, entropy, local repetition and state variables on held-out folios.

84. A generator that must be retuned for every section is weak

Unless independent evidence already establishes section-specific mechanisms, retuning after failure becomes patching.

85. The cross-transcription examination

Local similarity and entropy effects should survive reasonable alternative transcriptions.

Recent work reports substantial robustness across multiple transcription systems for major structural effects.

86. That strengthens the existence of the structure

It does not automatically strengthen self-citation as the cause.

87. The cross-section examination

A mechanism trained on herbal pages should predict what remains invariant in biological, pharmaceutical and stars sections.

88. Shared chassis plus regime state is the strongest version

The most defensible self-citation model is no longer “one simple rule makes everything.”

It is “a shared local-copy chassis operates inside changing structural regimes.”

89. This version explains more

It can accommodate local families, drift, page variation and section-level differences.

90. It also costs more

Regimes, slot grammar and source-selection rules all add parameters or latent structure.

91. Minimum description length becomes the judge

The stronger hybrid wins only if the extra machinery buys substantially better predictive fit.

92. The strongest rival is not random language

The correct rival is a structured meaningful system: technical language, abbreviation, notation or cipher with comparable surface constraints.

93. Straw-man controls would exaggerate self-citation

If the only alternatives are modern prose and random gibberish, self-copying will look uniquely successful.

94. Technical comparators are essential

Recipe books, catalogues, abbreviation-heavy manuscripts and structured records can reproduce many non-prose signatures while remaining meaningful.

95. The 2023 discussion literature matters here

Later discussion in Cryptologia has treated self-citation as one serious text-creation hypothesis among several rather than a settled solution. That is the appropriate current scientific posture.

96. CASE verdict dimension one: token families

Strong. Self-citation provides a simple, human-executable explanation for dense local families and near-neighbour structure.

97. Verdict dimension two: local repetition

Strong to moderate. A local source pool predicts repetition naturally, though semantic topicality can produce similar effects.

98. Verdict dimension three: frequency distribution

Compatible but weakly discriminating.

99. Verdict dimension four: entropy

Moderate. Structured self-copying can reproduce low-entropy behaviour, but other constrained systems can too.

100. Verdict dimension five: line effects

Conditional. The model needs explicit line-position or spatial copying rules.

101. Verdict dimension six: page states

Conditional. Drift or regime variables can explain them, but simple local copying is insufficient.

102. Verdict dimension seven: Currier structure

Partial. A shared copying chassis with changing regimes remains plausible; pure undirected drift is too simple.

103. Verdict dimension eight: cross-scribe behaviour

Compatible. A transmitted mechanism can operate across hands, but the evidence points toward regime structure rather than simple individual scribal style.

104. Verdict dimension nine: coherent word templates

Weak for simple self-copying. The most coherent regimes appear to require template structure beyond free local mutation.

105. Verdict dimension ten: image-text coupling

Unresolved. Seeding and label mechanisms are possible but not established.

106. Verdict dimension eleven: historical executability

Strong. The basic mechanism is feasible for a medieval scribe.

107. Verdict dimension twelve: historical purpose

Weakly constrained. Copying explains how, not why.

108. Verdict dimension thirteen: meaninglessness

Not established. Surface generation does not by itself determine semantic content.

109. Verdict dimension fourteen: complete explanation

Not yet. Important residual structure remains.

110. The overall verdict

Self-citation is one of the strongest partial mechanisms for explaining how Voynich word families and local repetition can arise. It is not currently sufficient, by itself, to explain the manuscript’s full structural hierarchy or establish that the text is meaningless.

111. This is not a rejection

The mechanism survives.

The strongest global interpretation does not yet survive with the same confidence.

112. CASE separates surviving core from failed extension

Surviving core: local copying and structured mutation plausibly contribute to surface organisation.
Needs extension: line, page, regime and template structure.
Unresolved: image-text relationship and semantic density.
Not established: complete meaningless-generation account.

113. The FAILURE Master does not retire the core

The correct state is not RETIRED.

It is CONDITIONAL / PARTIAL.

114. The SEARCH Master now gets better questions

How sharply does token similarity decay with distance?
Does vertical proximity outperform horizontal proximity?
Which mutation classes preserve legal templates?
Can one frozen copy-plus-template model reproduce line and page states?
Do label-to-prose transitions behave like seeded copying?
Can a meaningful technical comparator match the same signatures with equal or lower complexity?

115. The best next experiment

Build a constrained online generator with explicit source-selection, mutation classes, line reset and minimal slot grammar.

Freeze it before testing held-out folios across multiple scribes and sections.

116. The strongest rival experiment

Construct matched technical-language and abbreviation controls with comparable token constraints.

Ask whether semantics can reproduce the same signatures at equal or lower model complexity.

117. The discriminating result would be transfer

A theory that predicts untouched sections without retuning gains disproportionate confidence.

118. The discriminating failure would be template collapse

If synthetic copying repeatedly produces malformed word families when moved across regimes, local drift is not enough.

119. The discriminating semantic clue would be independent image-text prediction

If token classes predict independently defined visual referent classes beyond regime effects, a purely meaningless model weakens.

120. The discriminating null clue would be full reproduction without semantics

If one compact, historically executable generator reproduces the full hierarchy—including labels and cross-section behaviour—without semantic inputs, the null account strengthens dramatically.

121. The case study changes how we speak

Instead of “Voynich is generated” or “Voynich is language,” we can say something narrower and more useful.

Local copying is a serious mechanism candidate for part of the surface structure.

122. Narrow statements are cumulative

They can survive even when larger theories change.

123. CASE becomes a general article type

This article demonstrates why case studies deserve a dedicated longform family.

One hypothesis is taken seriously enough to earn its strongest form, then forced through the whole stack.

124. The standard CASE anatomy

claim;
mechanism;
best evidence;
cheap predictions;
expensive predictions;
strongest rival;
counterfactuals;
hold-out tests;
failure gates;
partial salvage;
verdict;
next experiment;
return to Masters.

125. CASE protects against straw men

The chosen hypothesis is presented in its strongest defensible form before critique.

126. CASE protects against theory worship

The article ends with a calibrated verdict rather than advocacy.

127. CASE protects against total rejection

Useful mechanisms are preserved even if a global interpretation fails.

128. CASE turns disagreement into architecture

Opposing camps can often agree on which local components survive even while disagreeing on semantics.

129. This is useful for eduKateAI

The AI does not need to choose one grand theory prematurely.

It can retain local mechanism confidence separately from global semantic confidence.

130. Typed edges from this case

EXPLAINS_PART_OF — local copying explains some surface structure;
REQUIRES_EXTENSION — additional mechanism needed;
SURVIVES_HOLDOUT — if future tests pass;
FAILS_AT — residual structural boundary;
DOES_NOT_IMPLY — local copying does not imply meaninglessness;
COMPETES_WITH — meaningful structured alternatives;
HANDOFF_TO — semantics, history, codicology or iconography;
RETURN_TO_KNOWLEDGE — only validated local claims promoted.

131. What KNOWLEDGE receives

Self-citation remains a serious published text-generation hypothesis.

Local copying can reproduce several important Voynich surface properties.

It is not established as the complete production history or proof of meaningless text.

132. What MECHANISM receives

Copying should remain one live component in mixed-model tournaments, especially at the token-family and local-context layers.

133. What MIRROR receives

Every copying signature needs a meaningful structured rival.

134. What SEARCH receives

Prioritise transfer, distance-decay, spatial directionality and template-coherence tests.

135. What FAILURE receives

Watch for unbounded grammar additions, page-specific retuning and semantic conclusions smuggled in from surface success.

136. The final case verdict

Keep self-citation in the machine. Remove it from the throne.

It is too useful to discard and too incomplete to crown.

137. That is what a mature verdict looks like

A mature research system is comfortable with partial truth.

It does not require every useful idea to become the whole answer.

138. The return to HUMAN

The receiver wants winners.

The evidence often gives components.

139. The return to the Warehouse

The case leaves behind a reusable set of tests, failure signatures, residuals and calibrated claims.

140. The return

The self-citation hypothesis entered this article as a candidate explanation for the Voynich Manuscript.

It leaves in a more precise state.

Local copying is probably one of the most useful mechanism classes to keep testing. Simple copying is not enough. Structured copying may explain more. Meaninglessness does not follow automatically. The decisive work now lies in transfer, template structure, image coupling and comparison against meaningful technical systems.

Primary routes from CASE STUDY

External research anchors

The operating rule

Keep the mechanism that survives. Retire only the claim that outruns it.

Next: Voynich Longform 08 · ADVERSARIAL / BREAK THE SYSTEM — deliberately attack the strongest surviving Voynich architecture with red-team tests designed to make it fail.