A Glyph Is Not a Letter, a Token Is Not a Word, a Space Is Not a Space: What the Units of Voynichese Are Not

Liudmila Rozanova; Alexander Temerev

cs.CLAug 17, 2026

A Glyph Is Not a Letter, a Token Is Not a Word, a Space Is Not a Space: What the Units of Voynichese Are Not

Authors: Liudmila Rozanova, Alexander Temerev

Organizations: International Institute for Applied Systems Analysis (IIASA), Laxenburg, Austria · University of Geneva, Geneva, Switzerland

Abstract

The Voynich manuscript (Beinecke MS 408) is usually analysed on three unstated assumptions: that its glyphs are letters, that the strings between blanks are words, and that every blank is a word space. We test all three against the Zandbergen-Landini transliteration with matched prose, cipher, and pseudo-text controls and quire-level resampling. None holds, and the failures share a shape: the order in Voynichese sits at the edges of tokens and at graded boundaries between them, not in the succession of tokens themselves. Glyph regularity is too strong for one-to-one substitution of any tested plaintext (conditional entropy 2.7 bits against about 3.5 for Latin, Italian, and English) and resolves instead onto a quire-stable scale of recurrent multi-symbol units. Tokens form a plausible vocabulary, yet the identity of one token predicts the next by under 1% of token entropy, below every matched control (2-10%), while the glyphs at token edges share 0.2 bits of mutual information, more than in any prose control. Blanks fall into two regimes: the separators transcribers marked uncertain behave like word-internal junctures, are physically narrower on the page (AUC 0.905 from independent image coordinates, with the same sign in a small blind ink audit), and are crossed by learned units even when every space is erased before learning. This profile is also what discriminates. A published Voynich-imitating cipher and a self-citation text generator both reproduce the low entropy, the unit scale, the weak token order, and the null result of a calibrated substitution attack; neither reproduces the edge-glyph coupling or the open, hapax-rich vocabulary (70% singleton types against 41% and 59-60%). Any account of the manuscript must therefore earn, rather than assume, the step from glyphs, tokens, and separators to letters, words, and word spaces, and these are the measurements on which to do so.

Figures & tables

Assumed equivalence	Main tests	Warranted interpretation
EVA glyph = plaintext letter	Conditional-entropy invariance; BPE scale; calibrated cipher attack	Fixed one-to-one letter mapping is rejected for the tested sources; recurrent multi-symbol units are the supported analytical scale.
ZL token = lexical word	Static vocabulary; shuffle-corrected token and edge-glyph order; uncertain-space merging	The forms look lexical in isolation but their identities do not carry the adjacent ordering expected of tested prose or catalogue words; the order that exists sits at the token edges.
Separator = uniform word boundary	Boundary association; image gaps; space-erased BPE	Blanks are non-random but form at least two graded boundary regimes; lexical status is not established.

Table 1: Three common unit identifications tested in this paper. Rejection as a default does not imply that no individual glyph, token, or separator can serve the proposed function.

Representation	Size in the main analysis	Uses
Composite-collapsed EVA	141,310 symbols; 47 observed types (25 frequent)	Boundary association, coordinate alignment, and the primary entropy comparison
Decomposed EVA	173,572 entropy symbols; 173,076 BPE symbols	Representation sensitivity and data-driven unit learning
ZL surface tokens	32,747 blank-delimited forms	Static vocabulary and adjacent-token order

Table 2: Observed representations used in the paper. “Symbol” is a transcriptional unit, not an assumed plaintext letter or palaeographic atom.

Corpus	Types	Singleton types	Top-50 coverage	Excess bits	Order share
Voynich, observed spaces	7,022	69.7%	33.2%	0.0661	0.79%
Voynich, weak spaces merged	7,920	73.0%	30.0%	0.0438	0.54%
Latin narrative	8,084	59.3%	26.7%	0.3256	4.00%
English narrative	4,248	51.7%	50.3%	0.7422	9.18%
French narrative	5,828	58.5%	44.5%	0.7828	9.72%
German narrative	6,920	63.4%	39.0%	0.3866	4.81%

Table 3: Static vocabulary and adjacent-token order. Continuous-text controls match the Voynich token count and line-length template. The catalogue matches token count while preserving record boundaries. Order values use a 2,000-type cap and subtract the mean of 100 within-line shuffles.

Figure 1: A word-like inventory with little prose-like order. (A) Distinct forms and coverage of the 50 most frequent forms at 32,747 tokens. Voynichese falls among the controls on these static measures. (B) Adjacent-token mutual information after subtracting the mean of 100 within-line shuffles, expressed as a share of capped token entropy. Continuous-text controls span narrative, medical, botanical, and herbal genres and match the Voynich line template; the catalogue retains record boundaries. Values use a 2,000-type cap.

	Last glyph $\to$ first glyph		Token $\to$ token		Edit-1
Corpus	bits	share	bits	share	ratio
Voynich, observed spaces	0.197	5.7%	0.067	0.81%	1.12
Voynich, weak spaces merged	0.162	4.8%	0.044	0.54%	1.14
Voynich, Currier A	0.137	3.9%	$-0.043$	$-0.50\%$	1.20
Voynich, Currier B	0.248	7.4%	0.029	0.35%	1.07
Latin botanical (lowest prose)	0.018	0.4%	0.148	2.03%	0.56

Table 4: Where adjacent order lives. Shuffle-corrected mutual information (100 within-line shuffles) between adjacent tokens, at the edge-glyph scale and at the token-identity scale (2,000-type cap), and the ratio of observed to shuffled frequency of adjacent tokens within one edit of each other. Full table in Appendix A.8 .

Boundary class	$I$	95% quire-bootstrap interval	$n$
Within-token pair	1.000	anchor	–
Uncertain separator ,	0.494	[0.464, 0.524]	2,409
First separator of line	0.099	[0.052, 0.145]	3,834
Mid-line separator	0.064	[0.002, 0.146]	21,836
Certain separator .	0.029	[-0.032, 0.111]	27,073
Continuation line break	-0.036	[-0.072, -0.002]	3,634

Table 5: Normalised internality by boundary class. Intervals directly resample the 16 quires and recompute the full estimator. Rows are overlapping cross-classifications, not one partition.

Figure 2: Boundary structure by transcription and by image. (A) Boundary association by transcription certainty and structural position; filled points are observed values with quire-bootstrap intervals, open points are expectations under the edge-composition null (not shown for the within-token anchor). (B) Bounding-box gaps by ZL certain/uncertain label; dashed lines show medians. (C) Mean gaps for the 12 most frequent flanking glyph pairs represented at both labels; blue denotes certain and red uncertain, as in panel B. (D) Mean box gap across equal-count bins of the transcription-only boundary association score; error bars are folio-bootstrap standard errors.

Figure 3: (A) Normalised internality index computed separately for Currier A and B. (B) Line-start enrichment of six glyphs, paragraph-first lines against all other lines, relative to the mid-line word-initial rate (log scale; connectors mark a sign reversal).

Corpus	$H(X)$	$H(X_{t+1}\mid X_{t})$	$2^{H(X_{t+1}\mid X_{t})}$
Voynichese, composite-collapsed EVA	3.98	2.69	6.45
Voynichese, decomposed EVA	3.89	2.32	4.98
Pliny, botanical Latin	4.01	3.51	11.39
Celsus and Caesar, Latin	3.99	3.47	11.10
Culpeper, English herbal	4.16	3.52	11.50
Italian prose	4.01	3.45	10.91

Table 6: Marginal symbol entropy, first-order conditional entropy, and conditional perplexity. All control corpora are truncated to the Voynich token count. Composite-collapsed EVA is the primary glyph-like representation; decomposed EVA is a sensitivity analysis.

Figure 4: Scale-dependent adjacent-unit structure under within-token BPE. (A) Dependence gap across BPE merge counts for Voynichese, Latin, two deterministic verbose substitutions, and a homophonic verbose substitution. All streams contain approximately 173,000 input symbols. The shaded band marks the 32–64-merge trough. (B) Currier A, Currier B, and the pooled corpus all reproduce a minimum at 64 merges. The orange dotted curve learns its merges outside each omitted quire and scores the held-out quire only; it reaches its minimum at 32. Stability annotations summarise quire-level refits; overlap is intersection size divided by the smaller learned-unit set in each disjoint-half partition.

Figure 5: Calibration and application of the learned-unit substitution attack. (A) Ground-truth class/letter association and attacked mapping recovery across five synthetic homophonic configurations of the same Latin; the wrong-language mapping is a contrast. (B) Every principal Voynich-minus-fitted-local-null language-match differential for Currier A and B; the solid vertical line is their mean. Shaded bands are the range of the same differential for the synthetic ciphers over five seeds each, against the right language (violet) and against a wrong language (grey).

Stream	$h_{2}$	$D_{64}$	Order	Hapax	Edge	Edit-1	Differential	Crossing
Voynichese	2.69	1.05	0.79%	70%	0.197	1.12	$-0.15$ to $+0.08$	2.5%
Naibbe cipher of Latin (Caesar)	2.71	1.03	1.25%	41%	0.006	0.82	$+0.00$	2.2%
Naibbe cipher of Latin (Pliny)	2.71	1.04	1.06%	42%	–	–	$-0.01$	1.9%
Self-citation generator (3 seeds)	2.6–2.7	0.65–0.86	0.5–0.6%	59–60%	0.006–0.009	1.06–1.23	$-0.02$ to $+0.02$	–
Latin narrative	3.44	1.24	4.00%	59%	0.033	0.43	(index 0.99)	15.5%
Synthetic homophonic ciphers	–	0.8–1.1	–	–	–	–	$+0.13$ to $+0.52$	3.8–12.4%

Table 7: The profile applied to two published accounts.

h_{2}

: first-order conditional entropy of composite-collapsed glyphs.

D_{64}

: BPE dependence gap at 64 merges (

D_{0}

is 1.56–1.63 for all Voynich-like rows and 0.52 for Latin). Order: shuffle-corrected token-identity share at the 2,000-type cap. Hapax: singleton share of types at 32,747 tokens. Edge: shuffle-corrected last-glyph

\to

first-glyph information. Edit-1: observed/shuffled ratio of adjacent tokens within one edit. Differential: language-match index minus that of the order-3 surrogate against the held-out Latin model (Voynich: range over 14 model comparisons; Naibbe: mean of five seeds; self-citation: range over three seeds). Crossing: share of hidden token boundaries crossed by learned units after space erasure (Voynich: certain separators). Details in Appendix A.5 .

Hidden position class	Crossed	Total	Crossing rate
Voynich certain separator	661	26,447	2.5%
Voynich uncertain separator	477	2,350	20.3%
Voynich all intra-line positions	97,402	162,330	60.0%
Latin word separator	3,796	24,464	15.5%
Deterministic verbose-2 word separator	1,528	12,360	12.4%
Homophonic verbose word separator	464	12,360	3.8%

Table 8: Crossing of hidden separator positions by units learned from space-erased line strings at 64 BPE merges. Control separators are genuine Latin plaintext word boundaries retained through encoding.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Blind direct-pixel validation (700-pixel Beinecke IIIF derivatives). (A) Direct ink-to-ink gap after blind QC: boxplot with the individual measurements (median line, mean marked by a white diamond); certain $n=265$ (mean 4.98), uncertain $n=21$ (mean 3.14), pooled certain $-$ uncertain $=+1.84$ px. (B) Mean gap by folio for the five folios retaining both label types; every folio shows the predicted sign (5/5; one-sided sign test $p=0.031$ ).

Figure 7: Robustness of the blind direct-pixel check. (A) Threshold stress test: certain $-$ uncertain mean gap across grayscale-threshold offsets around the bounded local Otsu value. (C) Objective geometry-concordance QC: the same contrast when only boundaries whose detected ink edges lie within a given tolerance of the locator-box edges are retained. (B) Alternative pixel estimators; the contrast stays positive for all but the vertical-overlap stress test.

Configuration	Right language	Wrong language
H2 (two homophones)	$+0.21$ [ $+0.09$ , $+0.25$ ]	$+0.09$ [ $+0.05$ , $+0.13$ ]
H3 (three homophones)	$+0.33$ [ $+0.11$ , $+0.44$ ]	$+0.09$ [ $+0.05$ , $+0.12$ ]
H5 (five homophones)	$+0.30$ [ $+0.23$ , $+0.34$ ]	$+0.10$ [ $+0.09$ , $+0.11$ ]
H3mix (one-to-three groups)	$+0.13$ [ $+0.09$ , $+0.22$ ]	$+0.06$ [ $+0.05$ , $+0.10$ ]
H3rot (rotating tables)	$+0.52$ [ $+0.45$ , $+0.61$ ]	$+0.10$ [ $+0.06$ , $+0.13$ ]
Voynichese (14 comparisons)	$-0.01$ [ $-0.15$ , $+0.08$ ]

Appendix

Table 9: Real-minus-surrogate language-match differential for the synthetic ciphers over five seeds (mean [min, max]).

Category	types	occ.	types	occ.	types	occ.
	$k=32$ (56 types)		$k=64$ (88 types)		$k=64$ , compound (63)
Single decomposed EVA letter (unmerged)	44.6	38.9	28.4	27.5	0.0	0.0
(i) one Zattera slot character	37.5	48.5	23.9	34.7	7.9	11.2
(ii) two slot characters, increasing slots	25.0	32.7	30.7	36.3	42.9	50.1
of which in adjacent slots	8.9	11.8	9.1	10.8	12.7	14.9
(ii ′ ) three or more slot characters	19.6	18.0	34.1	28.0	47.6	38.7

Appendix

Table 10: Learned BPE units against published word-structure inventories: shares (%) of unit types and of unit occurrences in the pooled decomposed-EVA stream at 32 and 64 merges; the last two columns restrict the 64-merge inventory to compound units.

Corpus	(a)	(b)	(c)	(d)	(e)	Destroyed	Edit-1
Voynich, observed spaces	0.197/5.7	0.243/4.3	0.199/5.8	0.167/2.0	0.067/0.8	90%	1.12
Voynich, weak spaces merged	0.162/4.8	0.181/3.3	0.149/4.4	0.121/1.5	0.044/0.5	88%	1.14
Voynich, observed, raw EVA	0.188/5.9	0.233/4.8	0.185/5.7	0.172/2.1	0.067/0.8	93%	1.13
Voynich, Currier A	0.137/3.9	0.134/2.4	0.081/2.3	0.045/0.5	$-0.043$ / $-0.5$	77%	1.20
Voynich, Currier B	0.248/7.4	0.280/5.2	0.211/6.3	0.191/2.3	0.029/0.3	91%	1.07
Latin narrative	0.033/0.8	0.220/3.4	0.174/4.3	0.268/3.3	0.326/4.0	78%	0.43

Appendix

Table 11: Adjacent-token information at five scales (bits/%).

Stream	Order share	Types	Hapax
Bencao Beiyao (herbal), tonal	5.40%	686	20.3%
Bencao Beiyao (herbal), toneless	5.23%	335	9.6%
Three Kingdoms (narrative), tonal	6.21%	749	17.6%
Three Kingdoms (narrative), toneless	6.42%	–	–
Voynichese, Currier A	$-0.53\%$	3,343	72.2%
Voynichese, Currier B (22,871 tokens)	$+0.41\%$	4,840	68.6%

Appendix

Table 12: Syllabic-transcription hypothesis against genre-matched pinyin controls at a matched token count. Order share is adjacent-token mutual information after the within-line shuffle correction, as a percentage of capped token entropy.

Mar 26, 2026

Evidence of Layered Positional and Directional Constraints in the Voynich Manuscript: Implications for Cipher-Like Structure

The Voynich Manuscript (VMS) exhibits a script of uncertain origin whose grapheme sequences have resisted linguistic analysis. We present a systematic analysis of its grapheme sequences, revealing two complementary structural layers: a character-level right-to-left optimization in word-internal sequences and a left-to-right dependency at word boundaries, a directional dissociation not observed in any of our four comparison languages (English, French, Hebrew, Arabic). We further evaluate two classes of structured generator against a four-signature joint criterion: a parametric slot-based generator and a Cardan grille implementing Rugg's (2004) gibberish hypothesis. Across their full tested parameter spaces, neither class reproduces all four signatures simultaneously. While these results do not rule out generator classes we have not tested, they provide the first quantitative benchmarks against which any future generative or cryptanalytic model of the VMS can be evaluated, and they suggest that the VMS exhibits cipher-like structural constraints that are difficult to reproduce from simple positional or frequency-based mechanisms alone.

Christophe Parisel

Structured Output Generation

Ask PDF

Oct 25, 2026

Does the Voynich Manuscript Mean Anything? Calibrated Tests Find Little Trace of Language

The Voynich manuscript is a fifteenth-century book in a script that no one can read. One account holds that its text means nothing and that a scribe made it by copying his own words and changing them a little. The other holds that a real language lies hidden behind a cipher. We write each account as a program that fills the pages of the manuscript, fit every program with the same budget, and check it on statistics the fit never saw. The copying process of Timm and Schinner predicts these statistics far better than any cipher with one key. But ciphers that switch keys between the manuscript's two dialects, or that encode only the word classes of a real text, predict them equally well. Word statistics alone cannot decide. We look for a trace of language that needs no decipherment. In 28 texts in 16 languages the same word pairs keep sharing a line. The manuscript shows this pattern at half the strength of the weakest of these texts. A word code keeps the pattern above the manuscript unless its plaintext has a weak pattern or many words are miscopied. Solvers that decipher control texts find no readable text in the manuscript under a letter cipher, a word code or 28 codes with one letter, letter pair or syllable per word. Word codes of abbreviated Latin, the plaintext their fits prefer, show which word follows which more clearly than the manuscript unless two words in five are miscopied. One cipher remains possible, a nomenclator with a second key whose scribe miscopies more than half of the words, and our solvers cannot exclude it in 9 of 16 languages. Most of the structure a solver finds comes from how the end of one word shapes the start of the next, which reflects how tightly glyphs are bound to each other. Planted messages are found reliably only when they are long, each letter always becomes the same word and these words are adjacent, so other messages could hide.

Xin Zhang, Victor S. Sheng

Texas Tech University, Lubbock, TX, USA

Topic pending

Ask PDF

Apr 28, 2026

A Quantitative Confirmation of the Currier Language Distinction

We present a unified quantitative analysis of the Currier A/B language distinction in the Voynich Manuscript, proceeding in two stages. First, we confirm that the distinction is genuine: a Beta-Binomial mixture model applied to character-pair substitution ratios across 185 folios, without access to Currier's labels, selects 2 by BIC and predicts held-out folio labels at 89% accuracy. Second, we show that the A/B contrast is not primitive but is a low-resolution projection of a higher-dimensional generative system. Its dominant component is a discrete boolean "switch" set once per folio, governing the vowel following the digraphs ch and sh. A two-state binomial mixture achieves Delta\ AIC = 2,549 over a single-state model and assigns 195 of 197 folios unambiguously. This switch does not operate uniformly: word templates divide into fixed contexts and switchable contexts, with template identity accounting for 92% of variance.

Christophe Parisel

Topic pending

Ask PDF

Abstract

Explore similar work

Evidence of Layered Positional and Directional Constraints in the Voynich Manuscript: Implications for Cipher-Like Structure

Does the Voynich Manuscript Mean Anything? Calibrated Tests Find Little Trace of Language

A Quantitative Confirmation of the Currier Language Distinction