Compared with published work

This examination finished its measurements before it checked the literature, so no earlier theory could steer them. It then listed every technique it had used, 211 in all, and searched for an earlier application of each one to the Voynich text. The search covered journals, the 2022 Voynich conference proceedings, arXiv preprints up to August 2026, René Zandbergen's site, and the main blogs and forums. Each technique was then sorted under deliberately conservative rules, which lean towards calling a technique established. A technique counts as established if someone had applied the same measurement to this text before, whatever controls were used here. It counts as adapted if this examination changed a published method in a named way. A named change is one of three things: a new null model, a new invariant or a validation step. A null model shows what the measurement gives by chance. An invariant is a quantity that some kind of encoding cannot change. A validation step runs the method on a text with a known answer. A technique counts as new only if the search found no earlier application at all.

The count depends on what is allowed to count as prior work, so the report gives it two ways. The first view counts every source the search found, including four unreviewed preprints from 2025 and 2026, student projects, code repositories and forum threads. Under that view, 32 techniques are new, 102 adapted and 77 established. The second view counts only peer-reviewed journals and the long-established sources, which are Currier, Tiltman, D'Imperio, Bennett, Stolfi and Zandbergen's site. Under that view, 55 are new, 103 adapted and 53 established. The two views disagree on 36 techniques, and each of those has a preprint, a student project or a forum thread as its only precedent. "New" means that the search found no earlier application. It does not mean that nobody has tried the technique.

The techniques that are new under both views.

• Sequential vs hierarchical within-word prediction test. This test measures how much the first glyph of a word tells about its later glyphs once the adjacent glyph is known. A nested category code, in which the first sign names a class and the later signs narrow it down, would keep that information, and a left-to-right template would not. The text keeps none of it, where Latin, Italian and Hebrew keep 0.3 to 0.6 bits.
• Hierarchical category-code generator. This generator tests the idea that the text is a philosophical language of the kind Wilkins proposed. In such a language a word is built by choosing a category, then a subcategory, and so on. The generator arranges the attested words in a branching tree (a trie) and makes a choice at every branch that is biased by frequency rank. It is fitted and scored on the full battery of statistics.
• Codebook/nomenclator generators. These control texts test the idea that each Voynich word is a code group standing for one word of a real text. Every word of a Latin, Italian or Finnish text is replaced by the Voynich word of the same frequency rank. The control therefore has no free parameter. Such texts fail on the statistics of similarity between neighbouring words.
• Segmentation-free repeated-substring test. This test removes all the spaces and counts the strings of 12 to 30 glyphs that occur more than once. The count does not depend on where the word breaks are. A fixed substitution cipher cannot change it, and neither can a verbose cipher, which writes each letter as a group of signs. The text has no repeated string of length 30, where the sixteen core texts have 28 to 3,899.
• Route / columnar transposition test. A transposition cipher keeps the signs and changes their order, for example by writing the message in columns. This test re-reads the glyph stream under nine page geometries (reversed lines, boustrophedon lines that alternate direction, columns, blocks and diagonals). It scores each reading on conditional entropy, which is how much surprise each glyph holds once the glyph before it is known, and on long repeats. Controls with a known key check that the test finds a transposition when there is one.
• Order-invariant phrase-repetition test. This test looks for repeated runs of words after each word is reduced to its bag of letters, with the order inside the word ignored. Anagramming the words cannot hide such repeats. Anagrammed Latin keeps 111 of them, and the text has 1 to 5.
• Codebook / nomenclator simulation grid. A codebook cipher replaces each word of a text with a code group. This grid builds 620 such texts from five plaintexts. It varies the homophones (several code groups for one word), the null groups (code groups that mean nothing) and the split words. Each text is scored together on the word-order information budget, the vocabulary statistics and the near-copy statistics.
• Labelling-free leading-unit distribution test. If the words were measurements written as numbers, their leading digits should follow Benford's law, under which small leading digits are commoner than large ones. This test compares the distribution of the leading units of the words with Benford's law and with a uniform distribution, in which every leading unit is equally common. It found no evidence either way, and nothing rests on it.
• Ordered-table tests. If the text were a table of numbers, neighbouring entries would often share their leading units and would often run in one direction. These tests measure such statistics of adjacent words (shared leading units, monotone runs) without needing to know what any sign means. They were calibrated on generated numeral tables (Roman, positional, ephemeris-like, sorted and accounting-like).
• Image-text association on the herbal pages. This test asks whether the words on a herbal page are linked to its picture. This examination computed descriptors from the scans of the herbal pages (pigment shares, plant shape, roots and flowers) and compared them with the page vocabulary by a Mantel test. That test asks whether pages with similar pictures have similar words. As a positive control, it planted a vocabulary linked to the pictures, which the test had to find.
• Fit-to-space at the line end measured on the scans. This measurement takes the space left at the right margin of 974 line ends on 42 pages from the scans. It compares that with a simulated copyist who fills the same lines with the same words and breaks the line when the next word will not fit. The text ends its lines within half a glyph of the margin and the copyist within one. The text is tighter on 13 of 15 pages with a clean margin.
• Ink change points against vocabulary drift. This examination measured the darkness and hue of the ink page by page from the scans. It then compared the places where the ink changes with the places where the vocabulary shifts. The two coincide no more often than chance would predict.
• Global shape clustering of aligned glyph crops. This examination cut small images of single glyphs from the scans without using a transcription. It grouped them by shape alone and compared the groups with the alphabet of the transcription. The grouping gave 30 clusters with a purity of 45 percent, purity being the share of each cluster that its commonest transcribed glyph makes up. Most of the mixing came from noise in the crops. Confidence that no forum experiment of this kind exists is low.
• Allograph-sequence channel test. Many glyphs have shape variants (allographs). This test asks whether the sequence of variants along the text is more predictable than a shuffled sequence. It allows for the scribe's hand, the page, the neighbouring glyph and the position. As a positive control, this examination replaced the variant labels with the vowels of a Latin text, to check that the test finds a hidden message when there is one.
• Residual-channel tests. A hidden message could live in the part of each word that the best predictive model cannot predict. These tests take that leftover as six streams (surprisal, rank, prefix, suffix, core and generating component). Each stream is tested for sequential structure, compressibility, repeats and position effects, against generators of the same shape. As a positive control, this examination planted a message of one letter per word, which the tests had to find.
• Word and glyph mutual information across line and paragraph boundaries. Across the line break the words share 0.008 bits of information, against 0.129 inside the line, and the glyphs share -0.003. Languages keep a third to two thirds of their in-line value across the break.
• Non-linguistic symbol corpora. The text was placed on a scorecard built for non-linguistic symbol systems. At the glyph level it scores a positional rigidity of 0.249 and an h2 of 2.07. Those values put it among the emblem systems and away from letters in words.
• Genuine historical ciphertexts. The Copiale cipher, a genuine 18th-century ciphertext, was measured as a reference class. Its h2 is 4.81 bits against 2.36 for the text, and it has no repeated run of 20 symbols.
• Word-pattern (isomorph) fingerprint. A word pattern records which positions in a word hold the same symbol, so no substitution cipher can change it. Only 15-16% of the merged-glyph tokens contain a repeated symbol, against 24-72% in alphabetic languages. The nearest scripts are the consonantal ones.
• Glyph-level homophone detection by hierarchical clustering of symbol contexts. The clustering of glyph contexts recovers artificial homophone keys exactly. On the text it finds no exchangeable pair of glyphs. The only candidates are variants that the standard transliteration already merges.
• Menzerath-Altmann law between word length in units and mean unit length. The Menzerath-Altmann law says that words with more units have shorter units. On minimum-description-length units the exponent is -0.59, inside the language range. The shuffled text and every generator reproduce it.
• Unsupervised word segmentation of the spaceless glyph stream and its agreement with the transcribed spaces. Word boundaries were recovered from the glyph stream with the spaces removed, at an F1 of 0.585 against 0.505 for Latin. The shuffled control recovers nothing, so the shape of the words is what marks the spaces.
• Long-context neural language model gain over Kneser-Ney n-grams. A character-level recurrent network was scored against n-gram models on the same folds. It beats the best Kneser-Ney model by 0.129 bits per glyph on the text, by 0.035 on the Markov imitation and by nothing on Latin. The gain lies inside the current word and the few words before it.
• Phonetic-prior decipherment into candidate known languages with a language-closeness score. The closeness score was calibrated with a positive control from the same language family. The text scores 0.05 or less against seven candidate languages. A Latin cipher scores 1.00 against Latin and 0.14 against its relative Italian.
• Clitic-stripping name test. The manuscript's labels are tested as running-text words minus a layer of proclitics. Stripping qo-, o-, y-, d- and s- from the running text moves its initial-glyph distribution away from the labels, by a divergence of 0.155 to 0.441 bits. It also halves the share of label types found in the text. The same operation on the Hebrew Bible brings text words to within 0.03 bits of the names that follow ben (son of), while leaving them attested in the text.
• Mixture and coined-noun plaintexts. Mixtures of Latin with Middle High German or Italian are constructed in four ways. They mix by sentence blocks, by word-level alternation, by replacing rare Latin words with vernacular ones, or by replacing them with coined strings. They are profiled at the manuscript's size. Every mixture raises the conditional letter entropy above both of its sources, to 3.36 to 3.43 bits against 3.11 to 3.30. Only word-by-word alternation removes the repeated four-word phrases, and it does so by destroying the adjacent-word information that the manuscript keeps. Coined vocabulary moves the one-edit family share away from the manuscript's value.
• Zodiac ring depth test. The twelve zodiac rings are tested for being one list of names written under twelve different alphabets. The largest set of labels that a single relettering maps from one ring onto another averages 3.6, against a chance level of 3.4 to 4.0. A planted list damaged by 20 percent gives 19.
• Core-only repeat counts. Repeated sequences are recounted on the word cores alone, after the prefixes and suffixes are removed by two grammars and by position. No reduction yields a repeated 25- or 30-glyph string. Latin with a null glyph at each word end recovers its 132 repeated 30-glyph strings when the ends are dropped.
• Leftover to the drawing outline. The space left before a drawing that interrupts a line is measured on the scans. On the herbal pages of the first hand the median is 0.8 glyph, against 2.7 for a copyist who breaks before the word that does not fit. Half the segments end within one glyph of the drawing.
• Word-gap profile along the line from full-resolution word boxes. Gaps between words and word widths are measured along 544 lines. The last gap narrows by about a tenth of a glyph and does not track the space left. The last word is 5 to 9 percent narrower than the line's mean. So the line end is fitted by the choice of word, and not by squeezing.
• Line-resolution ink runs. The darkness of the ink is measured line by line down each page of the manuscript. It is persistent, with a lag-1 autocorrelation of 0.44, and three quarters of the pages are individually beyond chance. It steps at 0.7 places per page, where a shuffled page steps at 0.02. On the pages aligned to the transcription, the steps fall on the first line of a paragraph three times more often than chance. The ink series records runs of the kind that dips, re-cuts and pauses produce. They are most frequent in hand 2 and rarest in hand 1.
• Energy-distance separation of generator clouds. Whole clouds of generator output, with each seed a vector of twenty-three text statistics, are compared with one another and with quire resamples of the text. The comparison uses the energy distance (Székely and Rizzo 2004) with a permutation reference. Every pair of generators separates at the smallest attainable p, and so does the text against every generator. The two halves of one generator's seeds do not.

The techniques that become new once the unreviewed work is set aside. These twenty-three have a preprint, a forum thread or a student project as their nearest precedent. The peer-reviewed literature has no precedent for them.

• Numbering-system test for labels. If the star and zodiac labels were numbers or names in a fixed order, consecutive labels on a page would resemble each other. They would be more alike than random pairs of labels from the same page. This test makes that comparison. Neighbouring labels are no more alike than random pairs. Forum posts have looked at adjacent labels, but none has used this test.
• Verbose/null-cipher solver. A verbose cipher writes one letter as a group of signs and may add nulls, signs that mean nothing. This solver searches at the same time for which symbols are nulls, which symbol pairs stand for one letter, and which letter each unit is. On Latin ciphers of that kind it recovered 91 to 97 percent of the words across the validation runs. On the text it behaves exactly as it does on the meaningless control. The nearest published work discusses verbose ciphers or builds one. It does not solve one.
• Held-out key-transfer test. This test fits a decipherment key on the Currier A pages and scores it on the Currier B pages, and the other way round. A genuine Latin cipher loses nothing in the transfer. The text loses 0.15 to 0.72 bits per character, which is what a meaningless control loses when its two halves are generated separately. Earlier work argued the same point from letter ratios and did not fit a key.
• Words-as-numerals test via slot dependence. This test measures how much the last two units of a word tell about its first two. The text gives 1.37 bits, where a positional number system gives nothing at all and Roman numerals give 0.2 bits. Generated numeral tables were the positive controls. The idea that the words are numbers had only been pursued by pattern matching before.
• Transcription-free glyph segmentation of full-resolution scans. This examination cut the full-resolution scans into lines, words and glyph-sized pieces without using a transcription. It then matched the pieces to the transcription by dynamic programming, a method that finds the best alignment of two sequences. Every crop then has a label and every line an alignment score. The only earlier work found is a recogniser trained on existing transliterations and described in blog posts.
• Paragraph-similarity network modularity against paragraph-shuffled text. The network of similar paragraphs was measured again with a matched reference set. Its modularity is 0.462 against 0.244 for the shuffle, inside the range of 28 languages in the same paragraph skeleton. The hand and Currier blocks explain 78% of the excess.
• Reading-direction cross-entropy asymmetry Delta_n = X_LTR - X_RTL. The reading-direction test was made symmetric in its treatment of the two directions. Held-out cross-entropy forward minus backward stayed under 0.001 bits per character for all 30 texts. The published figure is 0.065.
• Boundary signatures. The four boundary signatures were measured again with controls. The text scores 3 of 4, and so does its meaningless Markov-2 imitation. The criterion cannot tell the two apart.
• Internality index of uncertain word separators and line breaks. The index measures how strongly the glyphs on either side of a separator are associated. Uncertain spaces keep 0.29-0.43 of the within-word association, certain spaces 0.09-0.14, and line breaks none.
• Physical gap width on the scans as a predictor of transcriber space certainty. The widths of the gaps between words were measured on the scans. Spaces the transcribers marked uncertain are half as wide as certain ones (0.55 against 1.10 glyph widths) and ten times as wide as a gap inside a word.
• Dependence gap across BPE merge levels D_k = H. The dependence curve over merge levels has its minimum of 1.053 bits at 64 merges, which matches the published 1.045. Every language has its minimum at zero merges.
• Quire bootstrap and leave-one-quire-out confidence intervals for the headline statistics. Every headline number was given a quire-level interval. Dropping any single quire leaves h2 at 2.10-2.15 and the repeated 4-grams at 0-1. The spread between quires is 2-3.5 times the page-level noise.
• Paired single-leg gallows on paragraph top lines. The pattern of paired gallows on top lines was tested and is absent. Two or more such words appear on 56.6% of first lines, fewer than the 61.5% that random placement predicts, and they lie left of centre.
• Stroke-level symbol view. The alphabet was recoded into strokes. In that view h2 is 1.68 bits against 1.56 for Latin, and permuting the stroke table moves it by 0.21, so the view depends on the table.
• Two-language mixture test from the deviation of sorted letter frequencies from a logarithmic law, and spectral portraits of two-letter distributions. A mixture model asks whether the pages come from two sources. Its gain of 0.019-0.021 bits is 5-10 times that of any single language. Yet the gain persists inside each hand, and a third source still pays, which a real mixture of two languages does not show.
• Hildegard of Bingen's Lingua Ignota scored as a type list. The one medieval invented vocabulary of known meaning is scored beside the manuscript's word types. It shares the narrow ending system and the uninformative first glyph. But it keeps a language-like glyph entropy, 3.0 bits against 2.1 to 2.5, and few one-edit neighbours, about 20 percent against 50 to 87. So the first-glyph test cannot separate the manuscript from a constructed vocabulary, while the entropy and the neighbour density can.
• Two-manuscript scribal calibration of the Currier A and B divergence. The difference between the manuscript's two writing regimes, Currier A and B, is calibrated against two manuscripts of one text and two charter collections transcribed as written. Two hands copying one text differ in their letter distributions by 0.011 to 0.012 bits, four to seven times their split-half floors. A and B differ by 0.0196 bits, within a factor of two of that. The 0.19-bit difference in conditional entropy between A and B is matched by contiguous halves of B alone.
• Labels against period name lists. The manuscript's labels are set against six real name and word lists of its period, at matched size with resampled windows. The lists are litany saints, martyrology names, a monastic name list, a freemen roll, a plant synonymy and a glossary. The labels have a list's type-token ratio and singleton share. But 54 percent of them begin with one glyph, where no list puts more than 16 percent on its commonest initial. And 37 percent of their types lie one edit from a commoner type, against 7 to 14 percent in the lists. So they have the vocabulary shape of a list and the form of a template.
• Drawing-interruption edge test. The line-start rule and the loss of glyph association are measured where the pen restarts after a drawing inside a line. On the herbal pages of the first hand, the word after the restart is three quarters of the way to a line-initial word. The glyph association across the restart keeps 15 to 20 percent of its in-line value, where a true line break keeps none.
• Bifolium variance of held-out surprisal. Per-page predictability is partitioned by the physical make-up of the book. Inside hand and section strata the bifolium explains 56 percent of the variance of the held-out surprisal per page, against 22 percent for pages shuffled among bifolia. None of three generators in the same page skeleton reproduces the effect. Of four real texts poured into it, the German and the English show 0.44 at p 0.005, the Latin 0.29 and the Italian 0.20.
• Max-T over half-sample vectors. The twenty-three comparisons between the text and each generator are given one family-wise tolerance. It is the maximum absolute z over two hundred page half-samples of the text, a critical value of 3.16 in place of 1.96. That lowers the drift generator's misfit count from twenty to fifteen of twenty-three and leaves the joint rejection intact.
• Multiplicity ledger of harvested p-values. All 445 p-values saved for the text across a 153-technique examination are corrected together, by Holm within and across the five tool families and by Benjamini-Hochberg. Every demoted effect is named against its threshold. The 203 raw discoveries become 185 under the false-discovery correction and 137 under the family-wise one. The resolution floor of permutation p-values is shown to make the strict family-wise count a lower bound.
• Line-length regression with a within-page permutation null. Whether the number of words on a line depends on which words they are is measured by a held-out regression. It predicts the word count from the line's mean word length, its place in the paragraph and a vocabulary term. It is run on the manuscript and on Latin and generator text poured into the same layout, each line given the manuscript's own number of words. Mean word length explains 5 percent of the variance in the manuscript, against under 1 percent in the poured texts, but the poured texts are a null by construction, since the pour fixes the number of words on a line whatever the words are. The paragraph position term explains about 14 percent in every text alike, poured or not. Word identities add a further 1.4 percent over a within-page permutation null. That is a fit-to-space coupling, and no poured text could show one, so this test has had no control that could fail. A pour that fills each line to the manuscript's own width could fail, and it has not been run.

The adapted techniques that matter most. Two of them, the verbose solver and the key-transfer test, count as new once the preprints are set aside.

• Bias-corrected adjacent-word MI with an interior-only shuffle. Rozanova and Temerev's null shuffles every word in the line. The null here keeps the first and last word of each line in place, because edge words are different kinds of words (once-only words are twice as common among them). With a whole-line shuffle the correction flips sign (-0.12 bits) on this data. The edge-fixed version is what separates the real text from every copy and template generator. Whether the whole-line version does the same is not reported in the preprint.
• Verbose/null-cipher solver. Rozanova and Temerev fix the segmentation first, by byte-pair encoding, and substitute afterwards. The solver here searches the segmentation, the nulls and the letters together, with a stated cost for each null, and it reports word-level recovery on verbose ciphers with a known key.
• Many-to-one (homophonic) solver. Rozanova and Temerev, and Kuykendall, fix the number of classes in advance. Here the annealer chooses how many symbols share a letter, under a description-length penalty, and the collapse of the unpenalised objective to 'iiii...' is documented.
• 24 generative models from six families scored against a fixed 23-statistic battery, with fitted/trained/predicted bookkeeping per statistic, page-half-sample CIs and z-scores. Timm and Schinner, Gaskell and Bowern, Parisel, and Rozanova and Temerev all compared a generator with a set of statistics. Here 24 models are scored on one fixed battery, and a ledger records for each statistic whether the model was fitted to it, trained on it or predicts it. Only genuine predictions count.
• Ordered-slot grammar learned by greedy set cover. Stolfi and Zattera built their slot grammars by hand. The grammar here is learned automatically. It is scored by a curve of rule budget against coverage, by the productivity gap and by the shuffled-word selectivity test, and the same scoring is applied to Latin, Italian and Hebrew.
• Held-out key-transfer test. Parisel tested the same idea, that one key cannot serve both A and B, from the switching of character ratios. Fitting a real decipherment key on A, scoring it on B, and comparing the loss with a genuine Latin cipher (0.000) and a Markov control was not found.
• Cipher transformations of Latin. Bowern and Lindemann, and Rozanova and Temerev, enciphered known text and measured h2. The joint criterion, low h2 together with no long repeats, and the position-keyed variant are additions here.

What goes beyond the 2025 and 2026 preprints

Four preprints posted between September 2025 and August 2026 cover part of the same ground. Parisel wrote three of them (arXiv 2509.10573, 2604.19762 and 2604.25979) and Rozanova and Temerev wrote the fourth (arXiv 2608.17096). None has been peer reviewed. This examination therefore treats them as unreviewed claims. It does not treat them as a benchmark. The first count above includes them as prior work and the second sets them aside. Everything said about the preprints here comes from their own texts, and this examination did not run anything taken from them. The table lists, area by area, what they report doing and what this examination does that they did not.

The preprints also did some things that this examination did not, so this section is not a superset of them. Rozanova and Temerev ran a blind ink audit and used a catalogue and a pinyin text as comparison texts. Parisel measured the switching of four letter pairs between Currier A and B and built a Cardan-grille generator in Parisel's own form. A Cardan grille is a card with holes in it, laid over a table so that the holes pick out the parts of a word. The table-and-grille generator tested here is Rugg's. Four of the preprints' measurements were re-run here with controls: the validation of uncertain spaces from glyph coordinates, the reading-direction test, the byte-pair dependence gap and the resampling by quire. The section on the data and the sections each measurement bears on say what came of it.

The findings, old and new. About thirty of the results in this report reproduce published ones. Among them is the low glyph predictability, which is the low entropy of a glyph given the one before it. So is its cause, the fixed order of glyphs inside a word. The Zipf-like vocabulary, with a few common words and a long tail of rare ones, was known, and so were the narrow spread of word lengths and Sukhotin's vowel classes. The line was known to be a functional unit, with words behaving differently at its two ends. The absence of repeated phrases was known, and so were the networks of near-copy words. Currier A and B were known to follow the scribal hand. Earlier work had ruled out simple substitution, in which each sign stands for one letter, and had found that verbose encodings, in which one letter becomes several signs, survive. It was known that meaningless generators reproduce most of the single statistics. Rare words were known to come in bursts within a page, and each section was known to have its own vocabulary. The modularity of the paragraph-similarity network, the two-state letter grammar and the long-range correlations between words had all been reported. So had the low glyph-level rigidity on the symbol-system scorecard and the width of the uncertain spaces on the scans.

The search did not find the following results anywhere else. With the spaces removed, the text has no repeated string of 30 glyphs. Once the adjacent glyph is known, the first glyph of a word tells nothing about its later glyphs, which counts against a nested category code. On word-order information measured against the edge-fixed null, a shuffle that leaves the first and last word of each line in place, the real text separates from every meaningless generator. A cipher key fitted on Currier A loses information when it is scored on Currier B, where a Latin cipher loses nothing. Rank-matched codebook texts fail on the statistics of similarity between neighbouring words. The word-order information lies at the boundary between words. The line-edge derivation rates were quantified.

The search did not find these results elsewhere either. One slowly drifting state reproduces the whole multi-scale clustering of the vocabulary. The one-edit variants of words show no genealogy along the codex, that is, no line of descent as the pages go on. The line ends are fitted to the margin. The held-out predictive model leaves 11 bits per word unpredicted and gains 0.02 bits from word order. The Naibbe cipher separates from the text on vocabulary and page structure. Its agreement with the text on entropy and repetition is in the preprint by Rozanova and Temerev, whose Naibbe numbers agree with the ones found here. The argument from missing repeats holds when spelling variation is allowed for. No information crosses a line break. The claim of paired gallows on the top lines of paragraphs fails. The recurrent model gains over the n-gram models, and the gain lies within the word and the few words before it. The calibrated closeness score picks out no candidate language.

The later comparisons add more results that the search did not find elsewhere. The unicity distance of each cipher family, which is the length of ciphertext needed to pin down its key, was set against the length of the text. The Book of Soyga, the Lingua Ignota and the Steganographia were scored beside the manuscript as texts of known status. The first-glyph test does not separate a constructed vocabulary from the text. The generators competed on held-out quire halves, and a page classifier tells every generator from the text. The entropy margin below the nearest languages was measured at a matched sign inventory. The list genres of the period were compared at matched size. There are no isomorphic repeats (two strings with the same pattern of repeated signs) at 25 and 30 glyphs under any relettering. The lines of the first hand fit the drawings by word choice, and the line rule was tested again where the pen restarts after a drawing. The bifolium, a sheet folded once to give four pages, is a unit of the writing on two tests. Within one hand and section the letter distribution steps more at a change of sheet than at the turn of a leaf, and none of seven poured and generated controls in the same pages shows that step. And a bifolium of one hand bound in a quire of another keeps its own hand's statistics, not its quire's. Every generator is rejected jointly, with a calibrated tolerance. The minimum detectable effect of each null test was measured.

Five points differ from published readings. The size of the word-order signal depends on the null model. Rozanova and Temerev find 0.066 bits with a whole-line shuffle, and this examination finds 0.113 bits with the edge words held fixed. The signature of copying is diffuse here, where Timm and Schinner's generator copies locally. Landini and Montemurro read the long-range glyph correlations as linguistic organisation. This examination attributes them to the word template plus repetition. Parisel's reading-direction asymmetry does not survive a convention that treats the two directions alike. Arutyunov and colleagues inferred a mixture of two languages from the letter frequencies. This examination reproduces that gain, but the gain persists inside each scribal hand. So it reads as variation from page to page. It does not read as two languages.