# Procedure Fingerprinting for the Voynich Text: Which Mechanical Procedures of c. 1380–1520 Could Produce It, and How Each Would Show

Prepared 2026-10-22. Scope follows the brief: assume the text (Beinecke MS 408, ~37,000 tokens, five scribes c. 1404–1438, text statistically uniform across scribes) was produced by a replicable procedure that five people executed identically or that one person executed and the others copied. The question is which candidate procedures exist in the period, what statistical fingerprint each leaves, and what tests separate them from one another and from free composition.

## 0. Conventions

Citation tags: **VERIFIED-FETCHED** means the page was fetched in this session and the details come from it; **VERIFIED-URL** means the item and its URL/DOI were seen in a search result or in a fetched page but the item itself was not read, and any content attributed to it is from prior knowledge and flagged; **UNVERIFIED** means not located in this session, details from memory.

Facts taken as established by the program and not re-surveyed: conditional character entropy (h2) far below 103 languages; rigid slot grammar; words differing by one glyph behave alike in context; zero repeated 4-word phrases; word order carries 0.39 bits/word (languages 0.83–4.06); no held-out bigram gain; scribe explains <1% of vocabulary once subject is held constant; one-to-one and homophonic substitution of Latin/Italian/German/French/Czech, periodic transposition, position-keyed alphabets, nulls and sorted anagrams are excluded; vowel-less Latin is the nearest miss.

Numbers from outside the program that are used repeatedly below (all VERIFIED-FETCHED): h2 ≈ 2.7 bits against ≈3.5 for Latin/Italian/English controls; one token predicts the next by ≈0.8% of token entropy (controls 2–10%); the last glyph of one token and the first glyph of the next share ≈0.197 bits, more than any prose control; 7,022 types with 69.7% singletons; a reproducible BPE "trough" at 32–64 merges (Rozanova & Temerev 2026, arXiv:2608.17096). Zattera's 12-slot grammar covers 86.6% of tokens as regular, 10.4% as two regular parts, 3.0% unstructured (Zattera 2022, CEUR Vol-3313 paper10).

## 1. Historical table-driven encoding procedures, c. 1380–1520

The period's cipher record is dominated by Italian chancery practice. The organising fact for this survey is that between 1401 and 1450 the techniques did not change in kind, only in scale: more homophones, more syllables, more nulls, larger nomenclators (Pelling, "Fifteenth Century Cryptography", ciphermysteries.com 2016-07-06, VERIFIED-FETCHED).

### 1.1 Gabriele de Lavinde's ledger (Vatican, 1379)

Mostly simple monoalphabetic ciphers; many keys include nulls ("shapes that code for nothing") and a nomenclator of typically a dozen shapes standing for whole words (Pelling 2016, VERIFIED-FETCHED). Table shape: one entry per letter plus roughly 12 word-entries plus a handful of nulls, so ≈40 symbols. Word boundaries: Simonetta's 1474 rules analyse words of one, two and three signs and compare word-final signs with expected Latin and Italian endings, so boundaries were visible in the ciphers he attacked; his editor notes words were often left unseparated more generally (Simonetta, ed. Perret 1890, Project Gutenberg #28227, VERIFIED-FETCHED; Bonavoglia, HistoCrypt, also states Simonetta's ciphers left word spaces visible and were MASC, VERIFIED-FETCHED).

Predicted statistics: a relabelling of plaintext. h2 equals the plaintext's (≈3.5) minus a small amount from nulls that are inserted at random (random nulls raise, not lower, conditional entropy); fragment inventory ≈ alphabet size; within-word and across-word dependency identical to plaintext; no variant interchangeability; phrase repetition identical to plaintext (Latin formulaic 4-grams recur). Nomenclator entries create a small set of very frequent short tokens with no internal structure. Already excluded by the program.

### 1.2 Mantua cipher alphabet (1401)

Base line maps the alphabet to its reverse (A→Z); each of the five vowels gets four shapes; Pelling argues the vowel homophones were meant to hide which signs are vowels (defeating Simonetta-style vowel isolation) rather than to flatten frequencies (Pelling 2016, VERIFIED-FETCHED; Bonavoglia agrees homophones and nulls are "documented from the early 1400s", VERIFIED-FETCHED). Table shape: ≈20 consonant entries + 20 vowel entries ≈ 40 symbols, per letter. Nulls: present in the family. Boundaries: kept.

Predicted statistics: homophones on vowels raise unigram and conditional entropy of the ciphertext above plaintext if chosen at random, and leave consonant bigram structure intact. Variant interchangeability is real but restricted to vowel positions and positionally free (any variant can appear anywhere a vowel appears). Exact phrase repetition is broken in proportion to the number of vowel homophones (a 4-word Latin phrase with 6 vowels has 4^6 = 4,096 ciphertext forms), so zero repeated 4-word phrases is actually consistent with this family; but the program's test should collapse variants before counting repeats, and the collapsed text would then show plaintext-level repetition. The decisive failure is h2: homophones push it up, not down to 2.7.

### 1.3 Tranchedino ledger (Milan, c. 1450–1455) and the Mantua 1450 key

Vowels three shapes each, consonants two shapes each (except x, z and the "9" shorthand), own shapes for doubled letters and for scribal shorthand (the Tironian "9", ℞, "p"), more nulls than 1401, a nomenclator of more than eighty entries, and, in the Mantua 1450 key, a section of arbitrary shapes for syllable groups (ab, ac, ad, af, ag, ...) (Pelling 2016, VERIFIED-FETCHED). Bonavoglia describes Venetian practice as "syllable-based nomenclators" where breaking enough syllables lets the rest fall (VERIFIED-FETCHED). Table shape: ≈60 letter symbols + ≈80 word entries + a syllabary of order 100 + nulls, so an inventory in the low hundreds; entries per letter, per syllable and per word mixed in one key. Boundaries: kept in the attested dispatches of this type (Simonetta rules). Table-level counts are secondhand through Pelling's reading of Meister's editions of the ledgers (the editions themselves UNVERIFIED in this session).

Predicted statistics: this is the most Voynich-adjacent attested family because it mixes units of three sizes. Fragment inventory: 150–300 distinct symbols. Within-word dependency: lower than plaintext because syllable signs collapse letter sequences into single symbols, and the choice among letter-vs-syllable encodings for the same string is a second source of variant interchangeability beyond homophones. Across-word dependency: plaintext word order preserved, so word-order gain and held-out bigram gain would be at plaintext levels once variants are collapsed. Repetition: nomenclator tokens for Come/Quando/Quanto/Non recur as single symbols with exact repeats. Hapax: inflated by variant combinatorics but bounded by the inventory. The signature that separates it from a verbose cipher is that symbol count is at most equal to plaintext letter count (syllable signs shrink the text), whereas Voynich tokens are multi-glyph and more numerous than a plausible plaintext letter count would be under one-glyph-per-letter (Timm's estimate for Naibbe: 57,000–76,000 plaintext letters → 38,000 tokens, VERIFIED-FETCHED forum post).

### 1.4 Simonetta's rules (Pavia, 4 July 1474)

Thirteen rules for breaking MASC ciphers of Latin and Italian with visible word divisions: count distinct word endings to decide Latin vs Italian (Italian words end in five vowels; Latin may end in l, m, n, r, s or a mute), a one-sign word is "a", repeated word-final signs are vowels or s/m/t, a sign always followed by one other sign is q+u, doubled signs inside four-sign words are l or s, common digraphs/trigraphs map to che etc.; he states that nulls, mixing Latin and Italian, two alphabets, and a single sign for q and u defeat the rules (Perret 1890 edition, VERIFIED-FETCHED; Ycart, arXiv:1211.6847, VERIFIED-FETCHED, notes Simonetta never mentions frequencies, and Alberti gives orderings but almost no counts). Relevance: the rules are a period-authentic list of the ciphertext regularities a designer of c. 1400–1480 knew how to hide, which is exactly the list a Voynich-era designer would try to destroy: word-final letter distribution, one-letter words, the q-u lock, doubled letters. Voynichese has a strong word-final distribution (y, n, l, r, m), a strong q-o lock (qo), and no doubled letters except e/i runs, so it does not look like a text designed against Simonetta.

### 1.5 Alberti's disk (De componendis cifris, 1466/1467)

Two concentric 24-cell rings; the fixed ring carries 20 uppercase plaintext letters plus numerals 1–4, the movable ring 24 lowercase mixed cipher letters. The four numerals superencipher a 336-phrase codebook; enciphering a numeral also signals a disk rotation; the lowercase letters facing the numerals act as nulls; the index is reset "after three or four words" or on an inserted capital (Wikipedia "Alberti cipher", VERIFIED-FETCHED; Ycart, VERIFIED-FETCHED). Pelling and the 2016 article argue it had no discernible impact before the sixteenth century (VERIFIED-FETCHED). Table shape: 24 × 24 alphabet matrix plus 336 code entries. Boundaries: not discussed in the fetched sources.

Predicted statistics: a polyalphabetic with aperiodic key changes flattens h1 and h2 toward log2(24) ≈ 4.6; no slot grammar; no variant interchangeability; phrase repetition destroyed; word-order gain destroyed. It is the opposite of Voynich on every program statistic, and the program has already excluded position-keyed alphabets.

### 1.6 Trithemius: Steganographia (written c. 1499–1500, printed 1606) and Polygraphia (1518)

Steganographia Books I–II hide a message in a cover letter by a rule of selection, e.g. every other letter of every other word, sometimes combined with monoalphabetic substitution; Book III's numeric tables reduce modulo 25 with numbers ≤25 as nulls and a reversed 22-letter alphabet (Reeds, Cryptologia 22(4), 1998, VERIFIED-FETCHED via Gwern-hosted PDF; Ernst 1996 VERIFIED-URL).

Polygraphia Books I–II ("Ave Maria" cipher): each plaintext letter is replaced by a Latin word drawn from the column of a substitution table; columns are used in turn so that the output reads like a prayer. Book III uses invented words that look like inflections of one stem, and the plaintext letter is recoverable only from the word's position in the table, not from any word component. Book IV places the plaintext letter in the second position of an invented word (Hermes, "Polygraphia III", CEUR Vol-3313 paper7, VERIFIED-FETCHED). The tabula recta (progressive shift, key = counter) is also Trithemius' (Wikipedia "Tabula recta", VERIFIED-FETCHED). Table shape, Book III: 132 columns, each a full alphabet of invented words (Hermes simulated "all 132 columns" and a 10-column subset). Boundaries: each output word is one plaintext letter; plaintext word boundaries are not marked unless the encoder marks them.

Hermes' measurements are the only published statistics of a word-for-letter procedure against Voynich (VERIFIED-FETCHED): with all 132 columns word entropy 10.74, h2 2.90, h1−h2 1.10; with 10 columns word entropy 6.52, h2 1.90, h1−h2 1.67 against Voynich A/B word entropy ≈9.9, h2 ≈2.0–2.2, h1−h2 ≈1.7–1.9 (natural texts h1−h2 ≈0.8–0.9). Minimal pairs per unit: Voynich ≈30–35, P.III-all ≈11, P.III-small ≈222; minimal triplets Voynich ≈1.7–2.3, P.III-small ≈47; immediate word repetitions Voynich ≈8.5–9.6, P.III-small ≈26. Word-length type distribution is not binomial and types/tokens do not separate as they do in natural language (shared with Voynich).

Predicted statistics for the family: token count equals plaintext letter count, so 37,000 tokens ≈ 6,000–7,000 Latin words. Token-order gain is the plaintext letter-order gain (h1−h2 of Latin ≈0.6–0.9 bits), which lands near the program's 0.39 only if columns cycle. Held-out bigram gain would be strong (Latin letter bigrams generalise across sheets), and 4-word phrase repetition would be massive (Latin letter 4-grams such as "ibus", "tion" recur constantly). Both contradict the program's findings. Form–function coupling has the wrong sign in Book III: words differing by one component encode different letters (Hermes' rule "similar words encode different letters"), whereas Voynich minimal pairs behave alike. The Ave Maria version produces real Latin words with Latin statistics and is trivially excluded.

### 1.7 Later procedures (post-1520, outside the dating window but relevant as mechanisms)

Bellaso 1553 added a keyword selecting the alphabet (the "Vigenère" cipher); Vigenère 1586 published autokey (Wikipedia "Tabula recta", VERIFIED-FETCHED). Cardano's grille, 1550: a mask with irregular rectangular openings through which letters, syllables or words of the secret are written, the gaps then filled with cover text, the grille usable in four orientations; the Wikipedia page ties it to Rugg's Voynich proposal (Wikipedia "Cardan grille", VERIFIED-FETCHED; the attribution to De subtilitate is UNVERIFIED in this session). Porta 1563 digraphic table: UNVERIFIED in this session (not covered by the fetched tabula recta page). All are later than the radiocarbon window; Pelling's historical objection to Rugg rests on this (Schmeh 2016 part 2, VERIFIED-FETCHED). Their fingerprints: keyed polyalphabetics flatten h2 and destroy repetition; a Cardano grille used as intended produces a cover text with cover-language statistics, so neither matches.

### 1.8 Medieval Latin abbreviation systems and Tironian notes

Cappelli's elements classify abbreviations into truncation, contraction, marks significant in themselves, marks significant in context, superscript letters and conventional signs; the system "flourished, especially in Italy, from the 10th through the 15th centuries" (Medievalists.net summary of Cappelli 1982, VERIFIED-FETCHED; the usual figure of ~14,000 dictionary entries is UNVERIFIED here). Tironian notes: ≈4,000–5,000 classical signs, expanded in monastic teaching to ≈13,000–14,000; signs represent prepositions, truncated words, contractions, syllables and inflections and can be compounded; use declined, was revived by interest in the 12th century, and persisted into the 16th–17th centuries (Wikipedia "Tironian notes", VERIFIED-FETCHED). Davis describes abbreviation use in the Voynich scribes as "entirely discretionary" and proposes [f] and [p] as abbreviations of [te] and [ke] (Davis, "Voynich Paleography", CEUR Vol-3313 keynote2, VERIFIED-FETCHED).

Predicted statistics: a heavily abbreviated or devowelled Latin keeps plaintext word order and formulaic repetition, so order gain, held-out bigram gain and 4-gram repetition stay at language levels; h2 drops (Bowern & Gaskell found vowel removal and phonemic-contrast collapse among the closest matches on 42 word-level metrics, CEUR Vol-3313 paper6, VERIFIED-FETCHED), which is why the program finds vowel-less Latin the nearest miss. A Tironian-style sign inventory of thousands would give a very high hapax rate at the sign level, and a sign-per-syllable system would give a syllable-stream profile: Rozanova & Temerev report Pinyin syllable streams carry 5–6% order share and lower hapax rates than Currier A, so Stolfi's syllable hypothesis "fares poorly" on their measures (VERIFIED-FETCHED).

### 1.9 Number codes, syllabaries, musical and astrological notation

Trithemius Book III numeric tables (above) and the Alberti 336-phrase code are the attested number codes of the period. The "Sumerian-style syllabary" proposal for Voynichese is in the Stolfi tradition and is treated above. No literature on musical or astrological notation as a text-generation procedure was located in this session beyond the general observation that such notations have small inventories and strong positional grammars (UNVERIFIED). A notation-driven text would show h2 low (small inventory), a rigid slot grammar (fixed field order), no phrase repetition only if the underlying record (chart positions, melodic lines) never repeats, and no scribe effect; it would fail form–function coupling unless neighbouring symbols denote neighbouring values. It is included in the table in Section A as a "fixed-field record" family.

### 1.10 Summary of mechanism → statistics

| Procedure | Inventory | Unit | Nulls | Boundaries kept | Variant interchangeability | Phrase repetition |
|---|---|---|---|---|---|---|
| Lavinde MASC + nomenclator (1379) | ≈40 | letter + ≈12 words | yes | yes | none | plaintext level |
| Mantua 1401 (vowel homophones) | ≈40 | letter | yes | yes | vowels only, position-free | reduced by 4^vowels |
| Tranchedino/Mantua 1450 | 150–300 | letter + syllable + word | many | yes | letters and syllables, position-free | nomenclator repeats; else reduced |
| Alberti disk (1467) | 24×24 + 336 | letter, aperiodic key | yes | n/a | none | destroyed |
| Trithemius Ave Maria (1518) | 24 × columns | one Latin word per letter | no | not marked | none | plaintext letter 4-grams |
| Polygraphia III | 132 × 24 invented words | one word per letter | no | not marked | negative (similar word = different letter) | plaintext letter 4-grams |
| Abbreviated/devowelled Latin | hundreds–thousands | sign per syllable/word part | no | yes | scribal choice of abbreviation | plaintext level |
| Verbose bigram cipher (Naibbe) | ≈138 unigram types + prefix×suffix | 1–2 letters per token | p/m padding | respaced | positive, position-free | destroyed |

## 2. Greshko 2025, the Naibbe cipher

Citation: Michael A. Greshko, "The Naibbe cipher: a substitution cipher that encrypts Latin and Italian as Voynich Manuscript-like ciphertext", Cryptologia, online 26 Nov 2025, DOI 10.1080/01611194.2025.2566408 (title/author/venue VERIFIED-FETCHED via Semantic Scholar API and sci.news; the Taylor & Francis full text could not be fetched). The mechanics below come from secondary sources: the author (forum handle "magnesium") and Torsten Timm on voynich.ninja thread 5081 (VERIFIED-FETCHED), the Schneier blog comment thread (VERIFIED-FETCHED), and Rozanova & Temerev's reimplementation (VERIFIED-FETCHED). Nothing below is from the paper itself.

Mechanics as reported. Plaintext Latin or Italian is respaced into unigrams and bigrams by a dice rule ("snake eyes"), giving about 47.2% unigram and 52.8% bigram tokens; two respacing schemes exist, Simplified and Standard, the latter slightly favouring bigrams. Six substitution tables (α, β1, β2, β3, and two more) are selected by drawing playing cards in a 5:2:2:2:1:1 ratio, so the primary table α is used ≈38.5% of the time; a 78-card tarocchi and a 52-card variant exist. A unigram is written as one of six whole-token types per table (≈138 unigram word types in all); a bigram is written as a prefix (first letter) plus a suffix (second letter), prefixes drawn mainly from Zattera type-1 affixes and suffixes from type-2, giving 18 options per letter (6 unigram forms, 6 prefix forms, 6 suffix forms) and 19,044 theoretical prefix–suffix combinations. The codes are mutually prefix-free so decoding is unambiguous: a token that matches a unigram type is read as a unigram, otherwise as a bigram; if a bigram token collides with a unigram type the encoder redraws. Line-initial p and line-final m are treated as decorative or padding nulls. The author states any random or non-random source with the right ratios would do, and proposes that scribes could cycle tables deterministically.

What it reproduces (Rozanova & Temerev, VERIFIED-FETCHED): conditional entropy, the BPE unit scale, boundary permeability, weak token order, and it defeats their learned-unit substitution attack (so it is a cipher class their attack cannot see through). Secondary reports add symbol frequencies, word lengths and positional behaviour of certain glyphs (The Debrief, VERIFIED-FETCHED) and the letter-order cycle bias (Dillinger on Schneier's blog, VERIFIED-FETCHED). The author reports an R² ≈ 0.9 on a frequency-rank comparison (Timm's reading, VERIFIED-FETCHED).

What it fails: edge-glyph coupling across spaces is essentially zero against 0.197 bits in the manuscript; singleton share 41% against 70% (Rozanova & Temerev); long-range correlations, line- and paragraph-position effects, and word-type coverage (≈29% of Naibbe types occur in the manuscript, ≈45% of manuscript types are generable), all conceded by the author on the forum (VERIFIED-FETCHED). Timm's further objections: tables are populated with manuscript words and ratios fitted to manuscript frequencies, so validation on the same data is circular; free prefix–suffix pairing gives more type diversity and less within-word predictability than the manuscript; α's "e" token would be the single most frequent token (≈4.8%), weakening the homophony; 18 options per letter exceeds attested fifteenth-century practice (2–4); the apparatus of cards plus redraws is burdensome. The author answers that it is a feasibility demonstration, not a historical reconstruction.

Fingerprint for the program: a Naibbe-class text has (i) no information across the space (the space is a respacing boundary that never cuts a code unit), (ii) hapax share capped by combinatorics of the affix tables, (iii) exact 4-word repeats essentially zero (random table choice), (iv) positive form–function coupling (the six forms of one letter are interchangeable), (v) no line-position effects unless added, (vi) perfect scribe consistency if all scribes use the same tables and the same random source, but a detectable table-usage ratio per scribe if scribes cycled tables by habit. Point (vi) is the lever: in a card-driven version the per-letter form choice is memoryless; in a human-driven version it will show cycling or avoidance (see Section 4.4 and C.1).

## 3. Generative "no plaintext" procedures

### 3.1 Rugg's table and grille (2004; Rugg & Taylor 2016)

Citations: Rugg, Cryptologia 2004, DOI link 10.1080/0161-110491892755 (VERIFIED-URL from Schmeh's article; the title "An Elegant Hoax? A Possible Solution to the Voynich Manuscript", Cryptologia 28(1), is from memory, UNVERIFIED); Rugg & Taylor, "Hoaxing statistical features of the Voynich Manuscript", Cryptologia 41(3), 247–268, DOI 10.1080/01611194.2016.1206753 (VERIFIED-FETCHED via Crossref). Mechanics from Rugg's own blog series ("Hoaxing the Voynich Manuscript, part 7", VERIFIED-FETCHED) and Schmeh's 2016 three-part review (parts 1 and 2 VERIFIED-FETCHED).

Mechanics. A table is laid out in sets of three columns (prefix, root, suffix), about forty rows deep for one page, with cells deliberately left empty so some words lack a prefix, root or suffix; the table is filled one category at a time with syllables whose frequencies are "similar to Voynichese but not exact replicas". A card with three holes shows one prefix, one root and one suffix cell; each grille position yields one word; for the next word the grille moves three columns right and is shifted up or down by an irregular amount (regular vertical patterns would produce detectable sequences). When the line ends the grille returns to the first column set a little further down; when the table is exhausted a second grille with a different hole pattern restarts at the top. Immediate repeats may be kept or avoided. Rugg estimates ≈1.5 hours per page and under ten weeks for 240 pages; assistants with separate tables would produce different "dialects", his proposed explanation of A and B. Rugg's own noted side effect: rare syllables cluster at line starts and, after wrap-around, near line ends.

Statistics claimed: Rugg & Taylor 2016 claim Zipf's law, the symmetric word-length distribution, clustering effects and uneven syllable distribution (per Schmeh part 2, VERIFIED-FETCHED). Critiques: Schinner 2007 found the method "as investigated so far does not explain all of the statistical text properties" and suggested filling the table by a "lottery algorithm"; Pelling objects that the grille is 1550, that no table-and-grille reproducing a specific paragraph has been shown, and that matching paragraph-initial, line-initial, line-final and A/B behaviour needs many added rules (Schmeh part 2, VERIFIED-FETCHED). Montemurro & Zanette 2013 state that "not a single piece of quantitative evidence" had been advanced that such methods reproduce the manuscript's long-range structure (PLoS ONE 8(6): e66344, VERIFIED-FETCHED).

Fingerprint. Three-slot words (the manuscript needs twelve slots with 2–3 glyphs in nine of them, Zattera); word identity is a function of (row, column-set, grille) so the vocabulary is table-blocked: a page's types come from its table and recur across pages only when tables are reused; form–function coupling arises only through table geometry (words sharing a root are near each other in the table and therefore in the output), not through interchangeability; no cross-space glyph information; no held-out bigram gain; near-zero order gain; no long-range correlations beyond table reuse; line-position effects at line starts only; exact repeats of 2–3 words possible when the grille returns to the same rows, 4-word repeats negligible. Scribe consistency is high if one master table is shared, which Rugg does not assume.

### 3.2 Timm's self-citation and the Timm & Schinner algorithm

Citations: Timm, "How the Voynich Manuscript was created", arXiv:1407.6639 (v1 24 Jul 2014, v3 9 Dec 2015, 96 pp.), VERIFIED-FETCHED; Timm & Schinner, "A possible generating algorithm of the Voynich manuscript", Cryptologia 44(1), 1–19, online 25 May 2019, DOI 10.1080/01611194.2019.1596999, VERIFIED-FETCHED via Crossref (content from forum summaries, VERIFIED-FETCHED, and from Timm & Schinner's 2021 review, lingbuzz 005939, VERIFIED-FETCHED).

Rules as stated in Timm 2014 (VERIFIED-FETCHED). Source selection: similar groups occur within three lines above or below, about twice as often vertically as side by side; the scribe preferred the last completed sheet (repeats on subsequent sheets 37 times vs 5 across front/back of a leaf); initial words of an empty page come from another page. Modification rules I–IX: (I) replace one or more glyphs by similar-shaped glyphs; (II) add glyphs, usually a prefix; (III) delete glyphs; (IV) join two groups ("ol" + "chedy" → "olchedy"); (V) split such a group; (VI) duplicate a group or subgroup; (VII) copy verbatim occasionally; (VIII) add a gallows to the first group of a paragraph (617 of 716 paragraphs); (IX) combine I–VI and VIII. Similarity classes: in↔iin↔iiin; e↔ee↔eee; ee↔ch; ch↔sh; k↔t↔p↔f; chk↔ckh↔eke; o↔a; o↔y; n↔r↔m; l↔r↔m; r↔s; s↔d; with a vowel-like set (o, a, e, i, y) and a consonant-like set, consonant-like groups rarely consecutive. No probabilities are given for choosing among rules; the 2019 paper adds that source words are chosen "at least from the same page" and with some probability from the previous line at the same writing position (forum quotations of pp. 10 and 19, VERIFIED-FETCHED). Timm's own caveat: every written word becomes a potential source and the scribe could vary the method at will; critics (Zandbergen) say the method is therefore not well enough formulated to be tested, and Timm replies that it is a human process, not a program (VERIFIED-FETCHED). Timm reports a parameter scan in which "about 10–20% of the parameter space" matches key features within 10% relative error (VERIFIED-FETCHED forum post).

Statistics it reproduces, per the authors (VERIFIED-FETCHED, 2021 review): binomial word-length distribution, both Zipf laws, the tendency of high-frequency tokens to have more similar types, and the link between frequency, similarity and spatial vicinity; from Timm 2014: 84.67% of word types form one edit-distance-1 network, longest path 21, "daiin" has 36 neighbours, repeated words cluster within three lines, 35 phrases of ≥3 words recur ≥3 times with only 5 keeping order and 24 of 35 using similar rather than identical words; second words of lines shorter than first in 48% of lines vs 32%; p ends lines 62% of the time. Per Rozanova & Temerev's port calibrated only to mean token length and vocabulary size (VERIFIED-FETCHED): reproduces h2 (2.57–2.70), the BPE trough at 64 merges, weak token order (0.5–0.6%), an excess of near-identical neighbouring tokens, and a null result on their cipher attack.

What it fails (VERIFIED-FETCHED): edge-glyph information 0.006–0.009 bits vs 0.197; singletons 59–60% vs 70%; the BPE trough too deep (0.65–0.86 vs ≈1.05); type counts vary 5,000–7,100 across random seeds while the manuscript's values are stable across quires and Currier languages; no morphological word-class structure; and Feaster's positional findings ([Sh] vs [ch], [qo] vs [o], [k] vs [t] pairs systematically differ in line position in 19–21 of 21–23 high-frequency pairs, with the argument that repeated arbitrary modifications would converge, not diverge, such distributions; CEUR Vol-3313 paper12, VERIFIED-FETCHED). MarcoP found the generator's line-position plots flat (forum, VERIFIED-FETCHED).

Fingerprint. The unique, testable prediction is spatial: every token should have a near-identical predecessor within a few lines above, at a rate far above what topic clustering produces in a cipher of real text. Form–function coupling is produced by construction (variants are made from neighbours and placed near them), so coupling should decay with distance on the page. Scribe consistency is the second lever: five people improvising by self-citation from their own pages would diverge in vocabulary unless they seeded from the same sheets, and the seed-copying would itself leave shared rare types across scribes concentrated at page starts.

### 3.3 Schinner 2007

Citation: Schinner, "The Voynich Manuscript: Evidence of the Hoax Hypothesis", Cryptologia 31(2), 95–107, 28 Mar 2007, DOI 10.1080/01611190601133539 (VERIFIED-FETCHED via Crossref and Semantic Scholar; abstract not available; content from secondary sources). Method as described by forum readers (VERIFIED-FETCHED): each character mapped to a 5-bit pattern, the text turned into a binary random walk (a 28-character line ≈140 steps), compared with lowercase unpunctuated plaintexts; the walk shows "anomalous" behaviour and long-range correlations that Schinner read as better fitting a stochastic process than a natural-language encoding (Tablet Magazine quotation of Schinner, VERIFIED-FETCHED). Zandbergen reports the EVA-based plot is compatible with plaintexts up to about 28 characters and diverges gradually afterwards, which he tentatively links to something happening at line transitions (Currier's line-as-unit); Stolfi objects that natural texts also have random-walk behaviour and long-range correlations (VERIFIED-FETCHED). Montemurro & Zanette cite Schinner's fluctuation analysis as evidence of structure compatible with a ciphered real language (VERIFIED-FETCHED), the opposite reading of the same result. Net: the 2007 paper is a diagnostic (scaling of fluctuations), not a generator; its one durable claim is that whatever produced the text has correlations at the scale of a line and beyond, which any simulator must be checked against.

### 3.4 Gaskell & Bowern 2022, human gibberish

Citation: "Gibberish after all? Voynichese is statistically similar to human-produced samples of meaningless text", CEUR Vol-3313 paper4, VERIFIED-FETCHED. 42 volunteers (mostly Yale students) filled three pages with fake text in a made-up language, pen, lowercase, no punctuation, no algorithm given. 42 statistics on 200-word excerpts averaged over 100 draws, against 75 meaningful texts (≈5.6 M words) and five Voynich transcriptions. Gibberish has lower conditional entropy, worse Zipf fit, positive word-length autocorrelation (natural text negative), more repeated and tripled characters and words; Voynich shares all of these. Voynich plausibly resembles gibberish on 41 of 42 metrics and meaningful text on 38 of 42; the one gibberish failure is character-placement bias within words (charbias_words_mean), i.e. the slot grammar. A random forest put all five transcriptions nearer gibberish, with low confidence. Samples are short, so topic structure and page-to-page behaviour are untested; 17 of 31 illustrated samples show word lengths chosen to fit around drawings; the authors propose self-citation and self-correction as mechanisms. Companion paper (Bowern & Gaskell, paper6, VERIFIED-FETCHED): 22 manipulations of real text; closest matches to Voynich are phonemic-contrast collapse, prefix replacement, vowel removal and letter sorting; bigraph encoding brings Latin h2 to Voynich levels; nulls in some patterns raise word-formation predictability; no manipulation matches across the board; the authors flag the circularity of selecting manipulations by resemblance.

Related: Boxer's comparison with Dee/Kelley's Enochian (Sloane MS 3188): Enochian's running hapax fraction stays above 50% while Voynich behaves like natural text, so Voynich lacks the glossolalia fingerprint; both share epizeuxis (Voynich ≈1% of adjacent pairs, Enochian ≈0.15%, languages far less) (CEUR Vol-3313 paper1, VERIFIED-FETCHED).

Fingerprint of unconstrained human gibberish: low h2, no stable slot grammar, strong immediate repetition and word-length autocorrelation, high running hapax, vocabulary that drifts (Bowern's students struggled after ≈100 words, Schmeh, VERIFIED-FETCHED), and, decisively, no mechanism for five people to converge on <1% scribe effect.

## 4. Process identification: telling mechanical generation from copying, dictation and composition

### 4.1 Scribal error typology

The stemmatological lexicon classes all variants by result (omission, addition, transposition, substitution) and treats errors as "generally" copying errors while allowing dictation as a source; homoeoteleuton is defined as omission by eye-skip between similar endings (Parvum Lexicon Stemmatologicum, "Error" and "Homoeoteleuton", citing West 1973, VERIFIED-FETCHED). A fuller cause-based typology (VERIFIED-FETCHED from a secondary summary page, geosurge.ai, whose scholarly attribution is thin): visual errors (eye-skip/homoeoarcton/homoeoteleuton, letterform and ligature confusion specific to the script, haplography) diagnose an exemplar; auditory errors (homophone substitution, phonetic spelling, word-boundary shifts) diagnose dictation; memorial errors (transposition toward familiar patterns, harmonisation to a better-known parallel, correction from memory) diagnose recall; dittography and haplography diagnose loss of place. Young 1965 (GRBS 6(4), 247–274, VERIFIED-FETCHED abstract) tabulates Pindar errors as uncial-reflecting (few), aural (common) and several types of omission. Reynolds & Wilson's Scribes and Scholars and Havet's Manuel de critique verbale are the standard treatments (UNVERIFIED in this session).

Application. In an unknown script only visual and mechanical classes are observable: glyph-confusion substitutions between similar shapes (a/o, r/s, e/ch, the gallows), eye-skip between similar words within a line or across lines, and dittography. Pelling's survey of apparent slips by scribe (saiin for daiin, adeeody for odeeody, oechedy for orchedy, lr for ar, ro for ry; six slips in three lines of f26r and six in the first twelve lines of f115r, both Scribe 2, against none on f41r by Scribe 5 and "basically OK" for Scribe 3 on the rest of f115r; 12 of 13 "ry" at line end) is the only scribe-level error census located (ciphermysteries.com 2021-06-19, VERIFIED-FETCHED); Timm's rebuttal that "dain" occurs 211 times and "sain" 68 times shows the core problem: without a plaintext, "error" and "variant" are not separable by inspection, only by distribution (whether the rare form sits where the common form's context predicts the common form).

### 4.2 Stemmatics and exemplar-to-copy quantification

Haverals & Kestemont aligned a Middle Dutch exemplar and its direct copy word by word (Needleman–Wunsch), split the alignment into 75 ≈1,000-word samples and tracked variation rate over the copy, finding low variation at the start, a sudden rise around the 8,000-word mark, a near fourfold rise in the third scribe's stretch of the exemplar, and a sharp drop where a second exemplar likely begins; McIntosh's three modes (literatim, translating to own habits, mixed) frame the result (arXiv:2210.14061, VERIFIED-FETCHED). Camps & Randon-Furling model manuscript traditions as birth–death processes fitted by grid search on tree-shape summaries (survival, LCA distance to root, out-degree, Shannon index) and list variant introduction as future work (CHR 2022, arXiv:2210.16577, VERIFIED-FETCHED; follow-up arXiv:2505.19246 VERIFIED-FETCHED abstract). The method that transfers is the alignment-and-rate design: if two Voynich passages are copies of one exemplar passage (or if a scribe's page is a copy of a lost sheet), variation between them should be sparse, visual, and non-random in position; if they are independent executions of a procedure, variation is dense and uniform.

### 4.3 Generated-text detection statistics

GLTR scores each token by its probability, its rank under a reference model, and the entropy of the model's prediction; sampled text sits in the head of the distribution, and human text uses words outside the top 100 about 2.4 times as often (Gehrmann, Strobelt & Rush, arXiv:1906.04043, VERIFIED-FETCHED). DetectGPT uses log-probability curvature: perturbing generated text lowers its log-probability more than perturbing human text (Mitchell et al., arXiv:2301.11305, VERIFIED-FETCHED). The transferable idea is not the LM but the design: fit a reference model of the candidate procedure, then compare the rank/probability histogram and the curvature under small perturbations of the observed text with those of the procedure's own samples. A table-driven procedure with a random source is "flat" (every admissible choice at a slot is near-equiprobable, so the rank histogram is uniform over admissible options); a human-driven procedure is "peaked" (habitual choices) and shows avoidance or cycling; copying from an exemplar inherits the exemplar's profile exactly plus a visual-error layer.

### 4.4 Homophone-choice behaviour as a human signature

The Zodiac-340 solution used suppressed homophone cycling, avoidance of repeated symbols within a 17-symbol window and "pivot" shapes as diagnostics of how the encoder worked (Oranchak, Blake & Van Eycke, arXiv:2403.17350, VERIFIED-URL; details from the program's cryptanalysis report). Lasry's French Wars of Religion letter has two to six homophones per letter and shows encryption errors such as a T-symbol for F (HistoCrypt, VERIFIED-FETCHED). Megyesi et al. build synthetic keys with two or three code elements for frequent letters and one for rare letters, citing Kahn 1996 for that norm, and exclude nulls and nomenclature from their synthetic keys (HistoCrypt, VERIFIED-FETCHED). No study was located that measures per-clerk homophone-choice bias on historical ciphertexts; the Zodiac diagnostics are the nearest instrument and they transfer directly to Voynich minimal-pair variants (see C.1).

### 4.5 Model comparison for symbolic sequences

ABC model choice: posterior model probabilities by rejection, multinomial logistic regression or neural nets on summary statistics, with cross-validation (cv4postpr) producing a confusion matrix of how often each true model is recovered, and posterior predictive checks to see whether a model can reproduce a held-back statistic (Csilléry, François & Blum, abc R package, arXiv:1106.2793, VERIFIED-FETCHED). The caution that ABC model choice on insufficient summaries can be inconsistent is Robert, Cornuet, Marin & Pillai, PNAS 2011 (UNVERIFIED in this session); simulation-based inference with learned summaries is reviewed by Cranmer, Brehmer & Louppe, PNAS 2020 (UNVERIFIED). For this program the practical rule is: run cross-validation on the simulators before touching the manuscript, so that the confusion matrix tells you which procedure pairs the chosen statistics cannot separate; only statistics on which simulators are separable count as evidence.

### 4.6 The Rozanova & Temerev diagnostic set

Their paper is in effect a procedure-fingerprinting study (VERIFIED-FETCHED): they calibrate self-citation and Naibbe to the manuscript and find both reproduce h2, the BPE scale and weak order, while both fail edge-glyph coupling and hapax richness, and self-citation also fails stability across seeds and trough depth. They state that a hoax test "would have to rest on the vocabulary tail, edge-glyph coupling, trough depth, and stability across sections", and they list transposition, null insertion and state-dependent encodings as untested families. Those four quantities are the current best separators and are adopted in Section C.

## 5. Scribe-level tests

### 5.1 What is known about the five scribes

Davis identified five scribes by digital palaeography (Archetype/DigiPal) in Manuscript Studies 5(1), 164–180, 2020, DOI 10.1353/mns.2020.0011 (VERIFIED-FETCHED via Crossref and the Penn repository abstract). Her 2022 keynote (CEUR Vol-3313 keynote2, VERIFIED-FETCHED) gives the assignments: Scribe 1 writes the Language A botanical bifolia, f57v, Quires 15 and 17 (outer bifolium) and the Quire 19 recipe foldouts; Scribe 2 writes Language B botanical bifolia, all of Quire 13 (balneological), six Rose foldout panels and the first twelve lines of f115r; Scribe 3 writes the Quire 8 bifolium ff58/65, the inner Quire 17 bifolium and nearly all of Quire 20; Scribe 4 writes Quires 9–12 (astronomical/zodiac) and the nine-segment Rose; Scribe 5 writes ff57r/66v, f66r and separate Quire 6 bifolia. Language A is Scribe 1 only; Language B is Scribes 2, 3, 4 and 5; no scribe writes both; Scribe 4 has low [qo] and [ed] bigram frequencies. In the botanical section scribes alternate by bifolium (Quire 4: ff28/29, 27/30, 25/32 Scribe 1; ff26/31 Scribe 2), which Davis takes as evidence the bifolia are bound out of order and as evidence of collaborative production. Glyph-level discriminators are letterform details of [k], [n] finials and [f]. Currier (VERIFIED-FETCHED, voynich.nu transcription of his 1976 paper) counted six to eight hands, stressed A/B were statistical conveniences, found word-final symbols restrict the following word's initial (strongest in Biological B), found no repeated sequence crossing a line break in ≈25,000 words, and found one symbol ending the last word of a line 85% of the time.

Computational confirmation: character bigram/trigram TF-IDF classifiers separate Scribes 1–3 (Scribes 4 and 5 dropped for size; 31 pages each) at up to 90.3% (SVM, trigrams), Scribe 1 most distinct, Scribes 2 and 3 confused with each other, and the decision-tree root is "ed"/"edy", the Currier A/B feature; pages f57v, f65v, f85r, f86v, f95r, f95v, f116v disagree with the palaeographic label (Farrugia, Layfield & van der Plas, CEUR Vol-3313 paper5, VERIFIED-FETCHED). Note the circularity: this confirms that Davis' scribes align with the A/B and section split, not that the scribes differ once section is held constant, which is the program's own finding (<1%). Edwards (Goodreads blog 2024, VERIFIED-FETCHED; book UNVERIFIED) reports per-scribe tokens/types/hapax of 11,278/3,655/2,655 (S1), 9,813/2,686/1,908 (S2), 13,082/3,932/2,767 (S3), 5,576/2,652/2,070 (S4), 937/646/527 (S5), hapax rates of 19–24% of tokens for S1–S3, glyph-frequency correlations grouping S1 with S4 and S2/S3/S5 together, and {am}, {oe} in every scribe's top ten.

### 5.2 What a cipher executed by five people would show, versus copying from one exemplar

No published study measures hand-specific homophone bias in a multi-clerk historical cipher (Section 4.4). The expectations follow from the mechanisms. If five scribes each enciphered text with the same tables and a chance device (Naibbe-style), variant choice is memoryless and identical across hands; scribe would explain nothing beyond letterforms. If they enciphered by habit, each scribe's preference among interchangeable variants (the six forms of a letter, or the a/o, r/s, k/t alternations) would be a stable per-scribe bias measurable within a section, and errors would be idiosyncratic and uncorrelated across hands. If they copied one exemplar, variant choice would be identical across hands at the token level (the exemplar fixes it), errors would be visual and would include shared errors wherever two hands copied the same exemplar passage (none is known to exist in the manuscript, which is why shared-error tests must use within-text near-duplicates), and the only scribe effect would be letterforms plus a scribe-specific visual-confusion rate. The program's <1% figure is consistent with the first and third cases and against the second. Davis' bifolium alternation makes the "shared exemplar, parallel copying" scenario physically concrete: two scribes working on alternate bifolia of the same quire from a common source is a standard scriptorium pattern, while two scribes improvising the same slot grammar on alternate bifolia to <1% vocabulary difference is not.

Feaster's positional findings cut across all three: [Sh]/[ch], [qo]/[o] and [k]/[t] variants differ in line and paragraph position in nearly every high-frequency pair (VERIFIED-FETCHED). A memoryless chance device cannot produce that; a habitual encoder could if the habit is position-driven; a copied exemplar produces it if the source had it. Rozanova & Temerev's finding that uncertain separators behave as word-internal junctures, are physically narrower (AUC 0.905) and are crossed by learned units far more often (20.3% vs 2.5%) (VERIFIED-FETCHED) is itself a scribal-execution statistic: a space-width distribution that is bimodal and glyph-conditioned is a writing-process trace that should be compared across scribes.

## A. Candidate procedures × predicted fingerprint on the program's statistics

Entries: match / fail / partial relative to the program's observed value, with the mechanism in a word. "Collapsed" means after mapping interchangeable variants to one class.

| Procedure | h2 very low | 12-slot grammar | Form–function coupling (+) | 4-gram phrase repeats = 0 | Order gain 0.39 | Held-out bigram gain ≈ 0 | Hapax ≈70% of types | Scribe effect <1% |
|---|---|---|---|---|---|---|---|---|
| P1 MASC + nomenclator + nulls | fail (≈3.5) | fail | fail (none) | fail (plaintext repeats) | fail (≥0.83) | fail (strong) | fail (language-like, lower) | match if one key |
| P2 Vowel-homophonic (Mantua 1401) | fail (rises) | fail | partial (vowel variants only) | match raw; fail collapsed | fail collapsed | fail collapsed | partial (inflated) | fail if habitual; match if copied |
| P3 Letter+syllable+word key (Tranchedino) | partial (syllable signs lower it) | fail | partial | match raw; fail collapsed | fail collapsed | fail collapsed | partial | as P2 |
| P4 Verbose bigram cipher (Naibbe) | match | match (built on it) | match | match | match | match | fail (41%) | match (cards) |
| P5 Word-for-letter, real words (Ave Maria) | fail (Latin words) | fail | fail | fail (letter 4-grams) | partial (0.6–0.9) | fail (strong) | fail | match |
| P6 Word-for-letter, invented columns (Polygraphia III) | match with few columns | partial (stem+inflection) | fail (sign negative) | fail | partial | fail | fail (222 minimal pairs/unit) | match |
| P7 Alberti-type polyalphabetic | fail (≈4.6) | fail | fail | match | match (≈0) | match (≈0) | fail (very high) | match |
| P8 Abbreviated / devowelled Latin | partial (nearest miss) | fail | fail | fail | fail | fail | partial | match if copied |
| P9 Fixed-field notation (musical/astrological record) | match | match | fail | match if record non-repeating | match-ish | match | fail unless fields huge | match |
| P10 Rugg table + grille | match | partial (3 slots) | partial (table geometry only) | match | match (≈0) | match | depends on table reuse; page-blocked | match if one table |
| P11 Self-citation (Timm & Schinner) | match (2.57–2.70) | match (emergent) | match (by construction, distance-decaying) | match for exact, fail for near-repeats | match (0.5–0.6%) | match | fail (59–60%); unstable across seeds | fail unless scribes seed from shared sheets |
| P12 Unconstrained human gibberish | match | fail (charbias_words) | partial | match | match | match | high running hapax (Enochian-like) | fail |
| P13 Free composition, natural language | fail | fail | fail | fail | fail | fail | fail | n/a |
| P14 Any of the above, then copied by five hands from one exemplar | inherits | inherits | inherits | inherits | inherits | inherits | inherits + visual-error types | match |

Reading the table: only P4, P9, P10, P11 and P14-over-them survive the first seven columns, and each of the survivors fails something the program has not yet measured (edge-glyph coupling, hapax tail, vocabulary stability, spatial clustering of variants). The scribe column separates P11 and P12 from the rest unless the scribes copied.

## B. Simulator specifications (Python)

Common frame: every simulator emits lines and pages with the manuscript's line-length and page-length distributions (tokens per line, lines per paragraph, paragraphs per page, by section), so that line-position statistics are comparable; every simulator takes a seed; every simulator is run per Currier language with parameters fitted separately; outputs are EVA-like strings over the Zattera slot alphabet so the same scorers run unchanged.

B1 Chancery key (P1–P3). Inputs: Latin and Italian plaintext of matched genre (herbal/recipe/astrological), length ≈ 60,000–80,000 letters. Parameters: homophones per vowel hv ∈ {1,3,4}, per consonant hc ∈ {1,2}, syllable-table size S ∈ {0, 50, 100, 150}, nomenclator size N ∈ {12, 80}, null rate ρ ∈ {0, 0.03, 0.08}, boundary policy ∈ {keep, drop, respace}. Randomness: homophone choice ∼ Uniform or ∼ Markov with cycling parameter c (probability of taking the next homophone in a fixed order) and avoidance window w. Fit: hv, hc, S, ρ to h1/h2; c and w are the "human vs device" knobs.

B2 Naibbe-class verbose cipher (P4). Inputs: same plaintext. Parameters: respacing rule (p_bigram ≈ 0.53, or dice rule), table weights (5:2:2:2:1:1 or free Dirichlet), unigram forms per letter (6), prefix and suffix forms per letter (6, 6) drawn from Zattera type-1/type-2 affix lists, collision policy (redraw), nulls p/m at line edges. Randomness: table choice memoryless (cards) or cyclic (α, β1, β2, β3) or habitual (Markov with stickiness s). Fit: affix lists and weights to token frequency-rank; stickiness s is the human knob. Extensions to test Rozanova & Temerev's untested families: respace so that a code unit may straddle the space (introduces edge-glyph MI), and a state-dependent variant where table choice depends on the previous output token's last glyph.

B3 Word-for-letter table (P5–P6). Inputs: plaintext. Parameters: number of columns K ∈ {1, 10, 132}, column selection ∈ {cyclic, random, per-line}, word construction ∈ {real Latin words, stem + inflection with 12-slot templates}, whether plaintext word boundaries are marked. Fit: K and construction to h2 and minimal-pair counts (Hermes' targets). Prediction to check: sign of form–function coupling.

B4 Rugg table + grille (P10). Inputs: none (no plaintext). Parameters: table rows R (≈40), column-sets C, empty-cell probability per column e_P, e_R, e_S, syllable inventories and their table-fill frequencies (fit to Voynich syllable frequencies), grille count G per table, grille step rule (right 3, vertical shift ∼ Uniform(−v, v)), table reuse across pages (new table every T pages), repeat policy. Randomness: cell filling, vertical shifts. Fit: inventories and e to word-length and slot statistics; T to cross-page vocabulary overlap.

B5 Self-citation (P11). Inputs: seed page (sample of Voynich words or random words from the slot grammar). Parameters: source window (lines above ≤ L, same-line probability q_same, previous-sheet probability q_sheet, same-column preference), rule probabilities p_I..p_VII, similarity classes (Timm's list), paragraph-initial gallows p_VIII = 0.86, restart probability, and Stolfi's two-parameter minimal version (restart, mutation) as the baseline. Fit: calibrate to mean token length and type count only (Rozanova & Temerev protocol), then test everything else. Variant: five independent agents each with their own pages but sharing an initial seed sheet, to test the scribe-consistency prediction.

B6 Fixed-field record (P9). Inputs: a synthetic register (e.g., planetary positions or a melodic line) with field alphabets of 2–3 symbols in 9 of 12 fields. Parameters: field alphabets, per-field entropy, field-skipping probability (to create short words), record repetition rate. This is the control for "slot grammar without a plaintext language".

B7 Copying layer (P14). Inputs: any B1–B6 output as exemplar. Parameters per scribe: visual-confusion matrix over glyph pairs (a/o, r/s, e/ch, gallows), confusion rate ε_s, eye-skip rate λ_s with a similarity-conditioned trigger (homoeoteleuton on shared word endings), dittography rate, letterform-only identity. Output: five copies of disjoint exemplar ranges following Davis' bifolium assignments.

B8 Gibberish (P12). Use Gaskell & Bowern's released samples (github.com/danielgaskell/voynich, VERIFIED-URL) as the empirical generator rather than a parametric one; a parametric fallback is B5 with a drifting similarity list.

## C. Additional statistics, with null and control

C1 Variant-choice process statistic. For every minimal pair or variant set of interchangeable forms (program's own form–function classes, e.g. a/o, r/s, k/t, qo-/o-, Sh-/ch-), take the sequence of variant labels in reading order and compute (a) the run-length and lag-k autocorrelation of the label, (b) the transition matrix between variants (cycling index: excess of fixed-order transitions), (c) the avoidance index: probability the same variant recurs within w tokens relative to independence. Null: labels shuffled within scribe and section. Controls: B2 with cards (expect independence), B2 with stickiness s (expect positive autocorrelation), B1 with cycling c (expect fixed-order excess), B5 (expect strong positive autocorrelation with distance decay because variants are copied from neighbours), B7 (expect whatever the exemplar had, identical across scribes). This is the Zodiac homophone-cycling diagnostic transposed to Voynichese. Report per scribe; a per-scribe difference in s or c with no vocabulary difference is the signature of five humans executing one table by habit.

C2 Nearest-prior-similar distance. For each token, the distance in tokens and in lines to the nearest earlier token at edit distance ≤1 (Timm's similarity classes) and the share of tokens whose nearest similar predecessor lies within 3 lines. Null: the same text with lines shuffled within page, and with pages shuffled within section. Controls: B5 (strong excess within 3 lines, vertical > horizontal), B2 and B1 (excess only from plaintext topic clustering, equal in all directions), B4 (excess along the grille path: same line and adjacent column sets, not vertical). Timm reports 85–91% of 7–8-occurrence words have a similar word within three lines; the question is the excess over the null, which has not been computed.

C3 Vocabulary stability and overlap. Per quire and per scribe, with token counts equalised by subsampling: type count, singleton share, and the Heaps exponent; the Jaccard overlap of the top-200 types between quires, between scribes within a section, and between sections within a scribe. Null: Heaps curve fitted to the whole text, and overlap of random equal-size samples from the whole. Controls: B5 (type counts vary ≈5,000–7,100 across seeds and across independent agents; overlap between agents low unless seeded together), B4 (overlap governed by table reuse period T: page-blocked vocabulary), B2/B1 (overlap governed by shared function-letter tokens, high and uniform), B7 (overlap equals the exemplar's). Rozanova & Temerev's stability finding is the seed of this test.

C4 Edge-glyph information profile. Mutual information between the glyph at position −j of token t and the glyph at position +k of token t+1 for j, k ∈ {1, 2, 3}, computed per scribe, per line position (line-internal vs line-final/initial pairs) and separately across certain and uncertain spaces. Null: tokens shuffled within line. Controls: B2 respaced within units (expect ≈0), B2 with straddling units (expect MI concentrated at j = k = 1), B5 (≈0.006–0.009), B1 with boundaries kept (plaintext word-edge MI, ≈ language level), B6 (depends on field order). The manuscript's 0.197 bits is the strongest currently known separator and it has not been decomposed by distance, scribe or space type.

C5 Anomaly typology and shared-error coincidence. Take Zattera's 3.0% unstructured tokens and the program's own slot-grammar residuals; for each, find the nearest regular type and classify the edit by glyph-confusion class (visual pairs a/o, r/s, e/ch, i-count, gallows) versus non-visual; compute per-scribe rates and the visual share. Then test coincidence: for near-duplicate passages (sequences of ≥3 similar tokens recurring, Timm's 35 phrases), count anomalies occurring at the same position in both copies. Null: anomaly positions permuted within scribe. Controls: B7 (visual share high, rate differs by scribe, coincidences at exemplar-error loci), B5 (anomalies are new variants, not visual, and uncorrelated), B2 cards (anomalies are collision redraws, rare). Pelling's f26r/f115r census is the pilot for this, and its weakness (no criterion for "error") is what the nearest-regular-type construction fixes.

## D. Ranked tests most likely to move the question

1. Simulator cross-validation before any manuscript comparison. Build B1–B8, generate ≥200 texts per procedure with parameters drawn from fitted ranges, compute the program's eight statistics plus C1–C5, and run ABC model choice with cross-validation (confusion matrix). This tells the program which pairs of procedures its statistics cannot separate; everything else is reported conditional on that matrix. Without it, any "match" is unfalsifiable in the way Timm's and Greshko's are criticised for.

2. Edge-glyph MI decomposition (C4) by distance, space type and scribe. It is the single statistic every current generator fails, it is cheap, and its shape distinguishes a respaced verbose cipher (MI only if units straddle spaces) from a plaintext with kept boundaries (language-level edge MI) from self-citation (≈0). If MI concentrates at j = k = 1 and is identical across scribes, the text is a unit stream cut by a respacing rule applied identically by all five hands.

3. Variant-choice process (C1) per scribe. If interchangeable-variant sequences are memoryless and identical across scribes, the choice was made by a device or by an exemplar, not by five habits; if they show per-scribe stickiness or cycling with no vocabulary difference, five people executed one table by hand. This is the only test that directly addresses "five people executed it" rather than "what it is".

4. Nearest-prior-similar distance (C2) against the line-shuffled null and against B2/B4 controls. Self-citation's unique prediction is a within-3-lines, vertical excess; a cipher's excess is isotropic and topic-driven; Rugg's follows the grille path. One number (excess ratio, vertical vs horizontal) separates the three surviving no-plaintext and cipher families.

5. Anomaly typology and shared-error coincidence (C5). A visual-confusion-dominated anomaly class with scribe-specific rates and positional coincidences in near-duplicate passages is the copying signature; its absence, with anomalies that look like fresh variants, favours in-line generation. Combined with C3 (vocabulary stability across scribes within section), this decides between "five executions" and "five copies", which the table in Section A shows is the axis on which P11 lives or dies.

## Sources (fetched or verified this session)

- Rozanova & Temerev 2026, arXiv:2608.17096 — https://arxiv.org/pdf/2608.17096 — VERIFIED-FETCHED
- Greshko 2025, Cryptologia, DOI 10.1080/01611194.2025.2566408 — metadata VERIFIED-FETCHED (Semantic Scholar API; sci.news); full text not retrievable
- voynich.ninja thread 5081 (Naibbe discussion, Timm/magnesium) — https://www.voynich.ninja/archive/index.php/thread-5081-2.html — VERIFIED-FETCHED
- Schneier blog, "Substitution Cipher Based on The Voynich Manuscript", Dec 2025 — VERIFIED-FETCHED
- Hermes 2022, CEUR Vol-3313 paper7 — https://ceur-ws.org/Vol-3313/paper7.pdf — VERIFIED-FETCHED
- Gaskell & Bowern 2022, paper4; Bowern & Gaskell 2022, paper6; Farrugia et al. 2022, paper5; Zattera 2022, paper10; Feaster 2022, paper12; Boxer 2022, paper1; Davis 2022 keynote2 — all https://ceur-ws.org/Vol-3313/ — VERIFIED-FETCHED
- Timm 2014/2015, arXiv:1407.6639 — VERIFIED-FETCHED
- Timm & Schinner 2019/2020, Cryptologia 44(1):1–19, DOI 10.1080/01611194.2019.1596999 — VERIFIED-FETCHED (Crossref); content via forum and review
- Timm & Schinner 2021 review of Bowern & Lindemann, Cryptologia, DOI 10.1080/01611194.2021.1911875 — https://lingbuzz.net/lingbuzz/005939/current.pdf — VERIFIED-FETCHED
- Bowern & Lindemann 2021, Annual Review of Linguistics 7:285–308, DOI 10.1146/annurev-linguistics-011619-030613 — VERIFIED-FETCHED (Crossref metadata only)
- Schinner 2007, Cryptologia 31(2):95–107, DOI 10.1080/01611190601133539 — VERIFIED-FETCHED (Crossref); content via secondary sources
- Rugg & Taylor 2016, Cryptologia 41(3):247–268, DOI 10.1080/01611194.2016.1206753 — VERIFIED-FETCHED (Crossref)
- Rugg 2004, Cryptologia, DOI 10.1080/0161-110491892755 — VERIFIED-URL; Rugg blog "Hoaxing the Voynich Manuscript, part 7" — VERIFIED-FETCHED; Schmeh 2016 parts 1–2 — VERIFIED-FETCHED
- Montemurro & Zanette 2013, PLoS ONE 8(6):e66344 — https://pmc.ncbi.nlm.nih.gov/articles/PMC3689824 — VERIFIED-FETCHED
- Davis 2020, Manuscript Studies 5(1):164–180, DOI 10.1353/mns.2020.0011 — VERIFIED-FETCHED (Crossref; Penn repository abstract)
- Pelling, "Fifteenth Century Cryptography" (2016) and "New paper: Fifteenth Century Cryptography Revisited" (2017), ciphermysteries.com — VERIFIED-FETCHED
- Pelling, "What mistakes did Scribe 2 never make?" (2021) — VERIFIED-FETCHED
- Simonetta, ed. Perret 1890, Project Gutenberg #28227 — VERIFIED-FETCHED
- Bonavoglia, "Venetian Cryptanalysis Treatises of the Renaissance", HistoCrypt — VERIFIED-FETCHED
- Ycart, arXiv:1211.6847 — VERIFIED-FETCHED
- Wikipedia: Alberti cipher; Tabula recta; Cardan grille; Tironian notes — VERIFIED-FETCHED
- Reeds 1998, Cryptologia 22(4) — VERIFIED-FETCHED (Gwern-hosted PDF)
- Medievalists.net summary of Cappelli 1982 — VERIFIED-FETCHED
- Currier 1976 (voynich.nu transcription) — VERIFIED-FETCHED; voynich.nu a5_synt — VERIFIED-FETCHED
- Edwards 2024, Goodreads blog "Voynich Reconsidered: scribes and languages" — VERIFIED-FETCHED
- Parvum Lexicon Stemmatologicum, "Error" — VERIFIED-FETCHED; Young 1965 GRBS abstract — VERIFIED-FETCHED; geosurge.ai scribal-errors page — VERIFIED-FETCHED (secondary)
- Haverals & Kestemont, arXiv:2210.14061 — VERIFIED-FETCHED; Camps & Randon-Furling, arXiv:2210.16577 — VERIFIED-FETCHED; Camps, Randon-Furling & Godreau, arXiv:2505.19246 — VERIFIED-FETCHED (abstract)
- Gehrmann et al., GLTR, arXiv:1906.04043 — VERIFIED-FETCHED; Mitchell et al., DetectGPT, arXiv:2301.11305 — VERIFIED-FETCHED
- Csilléry et al., abc R package, arXiv:1106.2793 — VERIFIED-FETCHED
- Lasry, HistoCrypt (French Wars of Religion letter); Megyesi et al., HistoCrypt (historical language models) — VERIFIED-FETCHED
- Oranchak, Blake & Van Eycke 2024, arXiv:2403.17350 — VERIFIED-URL
- UNVERIFIED this session: Rugg 2004 title; Meister's editions of the Milanese ledgers; Reynolds & Wilson; Havet; Robert et al. 2011 PNAS; Cranmer et al. 2020 PNAS; Porta 1563; Cardano De subtilitate attribution; Cappelli entry count.