# Cryptanalytic and Decipherment Techniques for the Voynich Manuscript: Methods Survey Prepared 2026-10-22 from the collected research material (searches and page fetches logged in `a246681138a58db4c.full.md`) plus six supplementary lookups. Scope follows the brief: an inventory of cryptanalytic and decipherment techniques, with emphasis on what has not been applied to the Voynich text (Beinecke MS 408) and what could be run on a ~37,000-token transliteration. ## 0. Conventions **Citation status.** Every citation carries one of three tags. - **VERIFIED-FETCHED**: the page (or a page quoting the metadata) was fetched and the details below come from it. - **VERIFIED-URL**: the title and a URL or DOI were seen in a search result, so the item exists at that address, but the page itself was not fetched. Where I describe the content of such an item, that description comes from prior knowledge and is flagged as such. - **UNVERIFIED**: not located in the material or the supplementary lookups. Details given are from memory and should be checked before use. **Corpus assumptions used throughout.** The "EVA transliteration" is taken to be a line-level file in basic EVA (Takahashi, or the Zandbergen-Landini ZL file), with page, line and Currier A/B metadata, labels separable from paragraph text. Reported token counts vary by file and by how uncertain spaces are handled: Zattera's merged concordance gives 31,317 tokens and 5,105 types (VERIFIED-FETCHED); the brief's 37,000 matches the Takahashi full file. Any technique below should be run on at least three unit hypotheses, because the choice of unit is itself unsettled: - U1: basic EVA characters; - U2: slot units (ch, sh, cth/ckh/cph/cfh, ee/eee, ii/iii, ai/aii/aiin treated as single units), following Zattera 2022 and Zandbergen's STA; - U3: whole tokens (space-delimited), with uncertain spaces either kept or merged. **Controls.** Every scorer or solver should be run on the same footing against: (C1) Voynich with tokens shuffled within page; (C2) Voynich with glyphs shuffled within token; (C3) a synthetic text from Timm's self-citation generator; (C4) a synthetic text from Greshko's Naibbe cipher applied to Latin; (C5) Latin or Italian plaintext of comparable size enciphered with a Tranchedino-style key (three homophones per vowel, two per consonant, nulls, ~80 nomenclator entries); (C6) the unenciphered plaintext. A method that cannot separate Voynich from C3 and C4 while separating C5 from C6 tells us nothing. **What the Voynich text already looks like (facts that any cipher hypothesis must reproduce).** These are the quantitative constraints most relevant to cipher-type testing, all from fetched sources: - Conditional (bigram) character entropy about 2.7 bits, versus about 3.5 bits for Latin, Italian and English (Rozanova & Temerev 2026, VERIFIED-FETCHED). This is the single most cited obstacle to a simple substitution. - Glyphs group into recurrent multi-symbol units; one token predicts the next by under 1% of token entropy, versus 2-10% in prose controls; glyphs at token edges share about 0.2 bits of mutual information across the space, more than any prose control; about 70% of word types are singletons (Rozanova & Temerev 2026). - Word types fit a 12-slot positional grammar covering 86.6% of tokens as "regular" and a further 10.4% as two regular parts; 9 of the 12 slots admit only 2-3 glyphs (Zattera 2022, VERIFIED-FETCHED). - Identically and similarly spelled words cluster within the same line or one or two lines above, far more than in Arabic, Latin, English or German controls (Timm 2016, VERIFIED-FETCHED). - Within-word structure shows right-to-left optimisation while word boundaries show left-to-right dependency, a split absent from English, French, Hebrew and Arabic (Parisel 2026, VERIFIED-FETCHED). - Word-level metrics (42 statistics on 200-word samples) are best matched by Latin texts with phonemic contrasts collapsed, vowels removed, prefixes replaced, or letters sorted within words; bigraph encodings bring Latin character entropy down to Voynich levels; nulls in some patterns raise word-formation predictability (Bowern & Gaskell 2022, VERIFIED-FETCHED). - Section-specific word structure (not vocabulary) classifies pages into sections with 92% accuracy from four short character sequences, suggesting at least five "languages" beyond Currier A/B (Zattera 2022). Any solver that assumes a 1:1 glyph-to-letter mapping is therefore attacking a model the data already reject. The useful techniques are those that handle many-to-one, one-to-many and segmentation jointly, or that diagnose the cipher family before solving. --- ## 1. Homophonic, verbose and nomenclator cipher solving ### 1.1 Bayesian inference with Dirichlet priors (Ravi & Knight 2011) - **Source.** Sujith Ravi and Kevin Knight, "Bayesian Inference for Zodiac and Other Homophonic Ciphers," ACL 2011, https://aclanthology.org/P11-1025/ — VERIFIED-URL. Content description from prior knowledge. - **Assumes.** A homophonic substitution of a known plaintext language; a character n-gram language model of that language; each cipher symbol maps to exactly one plaintext letter; no transposition. - **Achieved.** First fully automatic solution of Zodiac-408 (from prior knowledge). Bayesian sparsity priors over the key, Gibbs sampling, and a stronger LM than earlier EM approaches. - **Tried on Voynich?** No published application located. The Voynich Ninja thread "Applicability of AZdecrypt to Voynich?" (Dec 2020, VERIFIED-FETCHED, replies not readable without login) shows the community recognised the homophonic assumption as the obstacle but no systematic run was reported. - **Implementation against EVA.** Sampling over a key from U2 slot-units (about 25-40 types) to a 23-letter Latin alphabet, scored by a 5-gram Latin LM. Because Voynich has fewer frequent units than Latin has letters, the plain homophonic direction (many cipher symbols to one letter) is the wrong direction; the model must be reversed or extended to allow one unit to carry a letter pair (see 1.9 and the verbose model in the ranked list). Run per Currier language and per section. Control: C5 must solve; C3/C4 must fail to produce a Latin LM score above the shuffled baseline. ### 1.2 Beam search decipherment (Nuhn, Schamper & Ney 2013; 2014) - **Source.** Malte Nuhn, Julian Schamper, Hermann Ney, "Beam Search for Solving Substitution Ciphers," ACL 2013, https://aclanthology.org/P13-1154/ — VERIFIED-URL. Same authors, "Improved Decipherment of Homophonic Ciphers," EMNLP 2014, https://aclanthology.org/D14-1184/ (also https://mirror.aclweb.org/emnlp2014/papers/pdf/EMNLP2014184.pdf) — VERIFIED-URL. Content description from prior knowledge. - **Assumes.** As 1.1. Key built incrementally symbol by symbol; partial hypotheses scored by an n-gram LM with heuristic future-cost estimates; beam pruning. - **Achieved.** Fast and accurate on 1:1 substitution; the 2014 extension handles homophones and solved Zodiac-408 automatically (from prior knowledge). - **Tried on Voynich?** Not located. - **Implementation.** Beam over U2 units ordered by frequency; expand each unit to a Latin letter (or a null, or a letter pair for verbose variants); score with a 6-gram LM; beam width 10^4-10^5. In Python on two cores this is feasible if the LM is a packed numpy array and the per-hypothesis rescoring is restricted to affected positions. Useful mainly as a cross-check on the annealing solvers of 1.3 and 1.7, since beam search gives a deterministic ranking of partial keys that can be inspected. ### 1.3 Simulated annealing with composite n-gram + entropy scoring, and the Zodiac-340 transposition workflow - **Source.** David Oranchak, Sam Blake, Jarl Van Eycke, "The Solution of the Zodiac Killer's 340-Character Cipher," arXiv:2403.17350 (2024), https://arxiv.org/abs/2403.17350 — VERIFIED-FETCHED. - **Method as fetched.** AZdecrypt (Van Eycke) uses specialised simulated annealing with up to 8-gram statistics (6-gram used in the batch runs). The score combines LM n-gram statistics with a plaintext entropy term that stops the solver converging on "n-gram spaghetti" built from a few common letters. Word divisions are inserted automatically from n-gram statistics. Transposition was detected by counting repeating bigrams at each period (spike at period 19, 37 repeats vs about 20 for shuffles), reinterpreted as a (1,2)-decimation of a 20x17 grid; about 655,088 transposition variants were sieved by repeat count and run through AZdecrypt, with ZKDecrypto runs on a supercomputer. Supporting diagnostics: suppressed homophone cycling, avoidance of repeated symbols within a 17-symbol window, "pivot" trigram shapes, unigram distribution, index of coincidence, and a deep-learning cipher classifier that flagged transposition. Verification: contemporary references, non-forced plaintext, systematic rule, FBI confirmation, and a unicity-distance argument. - **Assumes.** Homophonic substitution, possibly composed with a regular transposition, of a known language. - **Tried on Voynich?** Only the 2020 forum question. No published run with controls. - **Implementation.** The scoring function is the reusable part: S = sum log P(6-gram) + lambda * H(plaintext unigram). Implement in numba; a 37,000-token text is ~190,000 U1 characters, so each full rescoring is cheap and swap-moves can be scored incrementally. The transposition diagnostics are even cheaper and are listed separately in section 4.2 because they have never been run on Voynich. ### 1.4 The Copiale cipher (Knight, Megyesi & Schaefer 2011) - **Source.** Kevin Knight, Beata Megyesi, Christiane Schaefer, "The Copiale Cipher," ACL Workshop on Building and Using Comparable Corpora, 2011, https://aclanthology.org/W11-1202/ — VERIFIED-URL; Stockholm University project page https://www.su.se/english/research/research-projects/decipherment-of-historical-manuscripts/the-copiale-cipher-1.688288 — VERIFIED-URL. Content from prior knowledge. - **Method (prior knowledge).** A 105-page 18th-century manuscript of about 75,000 characters mixing Roman letters with abstract symbols. The first hypothesis (Roman letters carry the plaintext) failed; the successful model treated the abstract symbols as a homophonic cipher of German and the Roman letters as spaces/nulls, solved by EM-style key search with a German LM and manual bootstrapping from recovered words. - **Relevance to Voynich.** The lesson is methodological: the first step was to decide which graphemes carry information and which are nulls, by testing competing partitions with a language model. That partition step is exactly what Voynich work lacks (see ranked item 4). - **Tried on Voynich?** No. - **Implementation.** For each candidate partition of U2 units into "carrier" and "null" classes (e.g. gallows as nulls; line-initial glyphs as nulls; the "in"/"iin" tails as nulls), strip the null class and run the solver of 1.3. Compare the best LM score across partitions, with C5 controls that include a known null class. ### 1.5 The DECRYPT project: DECODE database, pipeline, and historical language models - **Sources.** - Beata Megyesi et al., "Decryption of historical manuscripts: the DECRYPT project," Cryptologia 2020, https://dare.uva.nl/id/9db9c73b-4f87-4cec-bea4-8a14f5b7d3c5 — VERIFIED-URL. - Mihaly Heder and Beata Megyesi, "The DECODE Database of Historical Ciphers and Keys: Version 2," HistoCrypt 2022, pp. 111-114, https://doi.org/10.3384/ecp188397 — VERIFIED-FETCHED. The database covers ciphers from the 14th century to 1965; the fetched page gives no record count beyond "tripled in size." - Heder, Fornes, Kopal, Szigeti, Megyesi, "Supporting Historical Cryptology: The Decrypt Pipeline," NEALT Proceedings Series 53, 2024, https://doi.org/10.58009/aere-perennius0100 — VERIFIED-FETCHED. Pipeline: cataloguing in DECODE, manual and AI-assisted transcription, cryptanalysis, historical-linguistic contextualisation. The authors report automated transcription is not yet accurate enough. - Megyesi, Sikora, Fornmark, Waldispuhl, Kopal, Mikhalev, "Historical Language Models in Cryptanalysis: Case Studies on English and German," HistoCrypt 2023, pp. 120-129, https://doi.org/10.3384/ecp195701 — VERIFIED-FETCHED. Hill climbing and simulated annealing with century-specific LMs; historical LMs give significantly better decipherment for ciphertexts from the 17th century or earlier, more so for longer and older texts. - Project pages: https://www.su.se/english/research/research-catalogue/research-projects/d/decipherment-of-historical-manuscripts — VERIFIED-URL. HistoCrypt 2025 proceedings: https://dspace.ut.ee/collections/47fa5504-4c9e-4cdd-8424-334344800f51 — VERIFIED-URL. - **Assumes.** The ciphertext belongs to a documented family (homophonic, nomenclator, code) whose keys resemble those in DECODE. - **Achieved.** A pipeline that has deciphered multiple early-modern homophonic and nomenclator documents; the 2023 result that LM period matters is directly actionable. - **Tried on Voynich?** The Voynich is in DECODE's period but no DECRYPT paper on it was located. The Voynich-conference report (Layfield et al. 2022, VERIFIED-FETCHED) is the closest overlap. - **Implementation.** Two concrete transfers. First, use DECODE's 15th-century keys as the generator for control C5 so that the synthetic cipher family is historically attested rather than invented. Second, replace generic Latin LMs with 14th-15th-century Latin and Italian LMs (Megyesi 2023's finding), including an LM trained on abbreviated (unexpanded) Latin, since a scribe enciphering from an abbreviated exemplar would produce different letter statistics than expanded Latin. ### 1.6 CrypTool 2 Homophonic Substitution Analyzer (Kopal) - **Sources.** Nils Kopal, "Solving Classical Ciphers with CrypTool 2," HistoCrypt 2018, https://ep.liu.se/ecp/149/010/ecp18149010.pdf — VERIFIED-URL (title inferred from query; page not fetched). "New Ciphers and Cryptanalysis Components in CrypTool 2," HistoCrypt, https://ecp.ep.liu.se/index.php/histocrypt/article/view/399 — VERIFIED-URL. Workshop handout "Breaking Homophonic Substitution Ciphers with CrypTool 2" (2019), https://www.cryptool.org/download/ct2/WS-2019_Breaking-Homophonic-Substitution-Ciphers-with-CrypTool-2.pdf — VERIFIED-FETCHED: semi-automatic mode with lock/unlock of letters, auto-locking of revealed words, a "Nomenclature" input for symbol-to-letter keys, and a note that "several of the cryptanalysis algorithms are based on implementations of George Lasry." The handout does not describe the internal algorithm. - **Assumes.** Homophonic or small-nomenclator cipher of a language for which CT2 ships n-gram tables. - **Tried on Voynich?** Not located. - **Implementation.** CT2 is a Windows GUI; its value here is as an off-the-shelf comparator for the Python solver on controls C5/C6, and for its ability to lock cribs interactively. The nomenclature input could be used to test a word-level hypothesis (e.g. that the most frequent Voynich tokens are nomenclator entries) by locking those and letting the solver attempt the remainder. ### 1.7 Lasry's search-metaheuristic methodology - **Sources.** George Lasry, *A Methodology for the Cryptanalysis of Classical Ciphers with Search Metaheuristics*, kassel university press 2018 (dissertation 2017), DOI 10.19211/KUP9783737604598, https://kobra.uni-kassel.de/handle/123456789/15274 — VERIFIED-FETCHED (abstract only; hill climbing and simulated annealing as the core, with genetic algorithms, ant colony and tabu search in the background). George Lasry, "Deciphering a Letter from the French Wars of Religion," HistoCrypt proceedings, https://ecp.ep.liu.se/index.php/histocrypt/article/download/402/360 — VERIFIED-FETCHED. Lasry, "Modern Cryptanalysis of Historical Ciphers" (CrypTool 20-years talk, 2019), https://www.cryptool.org/media/post-contents/20-years-cryptool/CT20years_Modern_Cryptanalysis_of_Historical_Ciphers.pdf — VERIFIED-FETCHED (lists Vatican 16th-18th-century homophonic and polyphonic ciphertexts attacked with simulated annealing; no Voynich mention). "Deciphering a Short Papal Cipher from 1721," https://ecp.ep.liu.se/index.php/histocrypt/article/view/401 — VERIFIED-URL. - **Method as fetched (French letter).** 858 symbols, 86 distinct, so not a simple substitution. Simulated annealing with a fitness score sum of count(trigram) x log(expected frequency). Moves: swap the plaintext letters of two homophones, or reassign one homophone; a cap on homophones per letter keeps the key balanced. Trigrams with a generic French corpus failed; 5-grams gave partial results; the break came with a corpus of historical French books followed by manual correction. The recovered key has two to six homophones per letter. - **Relevance.** This is the best-documented recipe for exactly the cipher family that 15th-century Italian chanceries used. Note the two failure modes it reports (too-short n-grams; wrong-period corpus) both apply to any naive Voynich run. - **Tried on Voynich?** Not located. - **Implementation.** As 1.3 but with Lasry's move set and homophone cap. The cap is important for Voynich because a solver without it will happily map 30 units to 5 letters to maximise n-gram score. Run with 5-gram and 6-gram Latin LMs built from period texts. ### 1.8 Dinu & Ciobanu - **Status.** UNVERIFIED. Two searches (one in the original material, one supplementary) returned only the DECRYPT project pages and a medievalists.net item on "Cryptographic Systems Used in the Romanian Countries between the 15th-19th Centuries" (https://www.medievalists.net/2015/02/cryptographic-systems-used-romanian-countries-15th-19th-centuries/, VERIFIED-URL, authorship not confirmed). No paper by Dinu and Ciobanu on cipher decipherment was located. Not used further. ### 1.9 Neural and LLM decipherment (2018-2026) - **Kambhatla, Bigvand & Sarkar, "Decipherment of Substitution Ciphers with Neural Language Models," EMNLP 2018**, https://aclanthology.org/D18-1102/ — URL seen in results; title match UNVERIFIED. From prior knowledge: beam search as in 1.2 but scored by a neural character LM. - **Aldarrab & May, "Can Sequence-to-Sequence Models Crack Substitution Ciphers?", ACL 2021**, arXiv:2012.15229, https://aclanthology.org/2021.acl-long.561 — VERIFIED-URL. From prior knowledge: a transformer trained end-to-end on synthetic 1:1 ciphers decodes unseen keys, with multilingual variants that also identify the plaintext language. - **"Segmenting Numerical Substitution Ciphers," arXiv:2205.12527 (2022)**, https://arxiv.org/abs/2205.12527v1 — VERIFIED-URL; authors (Aldarrab & May) from prior knowledge. The task: ciphertexts written as unsegmented digit strings where each letter is one or more digits, requiring joint segmentation and substitution. This is the closest published analogue to a verbose cipher. - **Bruton, Beloucif & Megyesi, "Attention-Augmented LSTMs for Automatic Homophonic Ciphertext Decipherment," arXiv:2606.05078 (June 2026)** — VERIFIED-FETCHED. Trained only on aligned ciphertext-plaintext pairs generated with ChronoFidelius from English and Swedish texts dated 1500-1899, with all keys drawn from one shared code pool. Near-perfect character accuracy, including short and noisy texts; fails predictably on codes outside the pool. - **"Benchmarking Large Language Models for Cryptanalysis and Side-Channel Vulnerabilities," arXiv:2505.24621** — VERIFIED-URL, not fetched. - **Assumes.** Supervised training on a synthetic cipher family; the test cipher belongs to that family. - **Tried on Voynich?** Not located. Turenne's arXiv:2609.20835 (VERIFIED-FETCHED) uses LLMs for image-text alignment and position-dependent probabilistic grammars, not decipherment. - **Implementation.** The honest use is as a family detector, not a decoder. Train a small seq2seq (or the Aldarrab-May segmenter) on Latin enciphered with (a) Tranchedino-style homophones, (b) Naibbe verbose tables, (c) Trithemian word-for-letter cover text; feed Voynich U2 sequences; measure whether the decoded output scores above the C1-C3 controls under a Latin LM. On a 2-core machine, training a 2-layer LSTM on ~1M synthetic characters is overnight work; a transformer is marginal. --- ## 2. Unknown-script decipherment ### 2.1 Snyder, Barzilay & Knight 2010 - **Source.** Benjamin Snyder, Regina Barzilay, Kevin Knight, "A Statistical Model for Lost Language Decipherment," ACL 2010, https://aclanthology.org/P10-1107 (PDF: https://pages.cs.wisc.edu/~bsnyder/papers/bsnyder_acl2010.pdf) — VERIFIED-URL. Content from prior knowledge. - **Assumes.** The lost language is related to a known language (Ugaritic-Hebrew); the script is alphabetic; words are segmented; a Bayesian model jointly learns a character mapping and cognate morpheme correspondences with sparsity priors. - **Achieved.** 29 of 30 Ugaritic letters and roughly 60% of cognates correctly mapped (prior knowledge). - **Tried on Voynich?** Not located. Hauer & Kondrak 2016 is the only comparable attempt (2.4). - **Implementation.** Requires a hypothesised related language and segmented words. Voynich U3 tokens are segmented, but whether they are words is disputed (Rozanova & Temerev 2026). The model can still be run with Latin, Italian, German, Hebrew and Arabic lexica as the "known" side; the interesting output is not the mapping but the posterior cognate rate, compared with C3/C4 controls where the true cognate rate is zero. ### 2.2 Luo, Cao & Barzilay 2019 - **Source.** Jiaming Luo, Yuan Cao, Regina Barzilay, "Neural Decipherment via Minimum-Cost Flow: From Ugaritic to Linear B," ACL 2019, arXiv:1906.06718, https://aclanthology.org/P19-1303/ — VERIFIED-URL. Content from prior knowledge. - **Assumes.** Known related language; character-level neural sequence model with a minimum-cost-flow constraint enforcing that each lost-language word maps to a distinct known word; cognate pairs are the training signal. - **Achieved.** Reproduced Ugaritic and recovered about two-thirds of Linear B cognates (prior knowledge). - **Tried on Voynich?** Not located. - **Implementation.** Same preconditions as 2.1; GPU not required at Voynich scale but training time in Python on two cores is hours per language hypothesis. The min-cost-flow constraint is valuable precisely because Voynich types are so numerous (70% singletons): it prevents the degenerate solution of mapping everything to a few Latin words. ### 2.3 Luo et al. 2021, phonetic priors and language identification - **Source.** Jiaming Luo, Frederik Hartmann, Enrico Santus, Regina Barzilay, Yuan Cao, "Deciphering Undersegmented Ancient Scripts Using Phonetic Prior," TACL 9 (2021), arXiv:2010.11054, https://aclanthology.org/2021.tacl-1.5 — VERIFIED-URL. Content from prior knowledge. - **Assumes.** Characters have sound values drawn from a universal phonetic inventory (IPA), so a prior over sound correspondences can be shared across languages; word boundaries may be missing (undersegmented); the lost language is one of a set of candidates. - **Achieved.** Gothic and Ugaritic decipherment without segmentation; a lost-language identification experiment on Iberian that ranked Basque among candidates without declaring a match (prior knowledge). - **Tried on Voynich?** Not located. This is the most directly applicable modern technique that nobody has reported running on Voynich. - **Implementation.** Treat U2 units as the undersegmented stream (ignore spaces, or keep them as soft boundaries); candidates: Latin, Italian, Occitan, German, Czech, Hebrew, Arabic, Greek, with lexica from any available wordlists. Output is a ranked language list with a confidence that is calibrated against C3 and C4 (where the correct answer is "none"). Note the Voynich-specific risk: if units encode letter pairs or abbreviations, the phonetic prior is misapplied; run it on U1, U2 and on a vowel-stripped Latin control to see what the method does when the assumption is violated. ### 2.4 Hauer & Kondrak 2016 and Zandbergen's critique - **Source.** Bradley Hauer and Grzegorz Kondrak, "Decoding Anagrammed Texts Written in an Unknown Language and Script," TACL 4 (2016), https://aclanthology.org/Q16-1006/ — VERIFIED-URL. Rene Zandbergen's review at https://voynich.nu/misc/misc_01.html — VERIFIED-FETCHED. - **Method and claims (as reviewed).** Language identification by character-frequency and character-pattern matching; monoalphabetic substitution solving; anagram decoding; vowel removal assumed. Hebrew ranked closest; a word-by-word translation of the first line was offered. - **Zandbergen's criticisms (fetched).** Only Currier-B pages (about 60%) in the Currier transliteration were used; frequency matching discriminates poorly (about 70% success on controls); Hebrew's 22 letters versus Currier's 35 symbols is unexplained; the translated first line is in Currier A, not B; mappings M/P and S/V are unresolved; the anagramming proposed cannot produce the observed low bigram entropy. He asks for repetition on FSG, EVA and v101 files. - **Later Kondrak-group work.** A University of Alberta repository item surfaced (https://era.library.ualberta.ca/items/1ff57e8b-825a-40bb-b1b6-941c76f97a3f, VERIFIED-URL, likely a thesis); nothing else was located. UNVERIFIED whether the group published further Voynich work. - **Implementation lesson.** Any language-ID experiment must be repeated across transliterations and across Currier A/B, with the same pipeline applied to C3/C4, and must report the discrimination rate on known-language controls at matched length. ### 2.5 Reddy & Knight 2011 - **Source.** Sravana Reddy and Kevin Knight, "What We Know About The Voynich Manuscript," ACL LaTeCH workshop 2011, pp. 78-86, https://aclanthology.org/W11-1511/ — VERIFIED-FETCHED (metadata only). Content from prior knowledge: a two-state HMM separates glyph classes suggestive of a vowel/consonant split; word-length distribution is unusually peaked; low bigram entropy; page-level similarity confirms Currier A/B; they raise the abjad possibility. - **Status.** Already applied to Voynich; listed here as the baseline that section 2.3 would extend. --- ## 3. Historical cipher practice c. 1400-1470 and Voynich-specific cipher models ### 3.1 Attested chancery ciphers (Lavinde 1379, Mantua 1401, Tranchedino c. 1450, Simonetta, Alberti 1467) - **Source.** Nick Pelling, "Fifteenth Century Cryptography," ciphermysteries.com, 6 July 2016, https://ciphermysteries.com/2016/07/06/fifteenth-century-cryptography — VERIFIED-FETCHED (first 100,000 characters). Pelling, "Fifteenth Century Cryptography Revisited" (18-page draft on academia.edu, announced 8 July 2017), https://ciphermysteries.com/2017/07/08/new-paper-fifteenth-century-cryptography — VERIFIED-FETCHED. Tranchedino facsimile listings: https://facsimilefinder.com/facsimiles/francesco-tranchedino-secret-diplomatic-documents-facsimile — VERIFIED-URL. Primary editions (Meister 1902; Tranchedino facsimile 1970; Cerioni; Costamagna 1968) are mentioned in the fetched post and its comments but were not themselves retrieved: UNVERIFIED as citations. - **Features as fetched.** - Lavinde's Vatican ledger (1379): mostly monoalphabetic, many with nulls, some with nomenclators. - Mantua secretary's alphabet for Simeone de Crema (1401): reversed-alphabet core, multiple shapes per vowel. - Tranchedino ledger, Milan (c. 1450-55): three shapes per vowel, shapes for doubled letters, shorthand abbreviations, more nulls, a nomenclator of more than eighty entries. - Tranchedino, Mantua ducal cipher (1450): two homophones per consonant (except x, z, "9"), three per vowel, mini-codebook of common words, nulls, arbitrary shapes for syllable groups. - Simonetta's treatise (Milan, 1450s-60s): advice on vowel identification and common-word cribs. - Alberti's *De Cifris* (1466/67): explains frequency analysis; Pelling argues it had no discernible impact before the 16th century. Pelling's 2017 draft argues homophones were a steganographic defence against linguistic weaknesses of Italian/Latin rather than a response to frequency analysis. - Wikipedia pages on the Alberti cipher and polyalphabetic ciphers appeared in results (VERIFIED-URL) but add nothing specific. - **Why this matters for statistics.** A Tranchedino-style key produces: more cipher symbols than plaintext letters (the opposite of Voynich's small unit set unless units are verbose), vowel homophones that flatten vowel frequency, nulls that raise entropy, and nomenclator tokens that produce a long tail of frequent fixed words. None of these reduces conditional entropy to 2.7 bits. The only attested features that could lower entropy are the syllable-group shapes (a form of verbose or polygraphic substitution) and abbreviation shapes. This is why the Voynich-specific models in 3.3-3.5 all add verbosity. - **Implementation.** Build the C5 control generator from these keys literally (three vowel homophones, two consonant homophones, nulls at a tunable rate, 80-entry nomenclator from the most frequent Latin words). Then measure the same 42 Bowern-Gaskell metrics plus conditional entropy. This gives a quantitative statement of how far attested 1450 practice is from Voynichese, which has not been published. ### 3.2 Trithemius and the "Polygraphia III" model (Hermes 2022) - **Source.** Jurgen Hermes, "Polygraphia III: The cipher that pretends to be an artificial language," International Conference on the Voynich Manuscript 2022, CEUR-WS Vol-3313 paper 7, https://ceur-ws.org/Vol-3313/paper7.pdf — VERIFIED-URL (programme entry VERIFIED-FETCHED). Content from prior knowledge: Trithemius's third book of the *Polygraphia* (1518) encodes each plaintext letter as an invented word whose form follows a regular morphological scheme, so the output looks like a constructed language with strongly regular word structure. - **Relevance.** A word-for-letter cipher is the one historically attested family that yields highly regular word-internal structure, low conditional entropy, and vocabulary that scales with the plaintext alphabet rather than with the plaintext lexicon. It postdates the Voynich radiocarbon window (early 15th century) but Trithemius drew on earlier practice. - **Tried on Voynich?** Hermes presented the comparison at VOY2022; no solver run was located. - **Implementation.** Under this model each Voynich token encodes one or two plaintext letters. Test: fit a token-to-letter homophonic solver (1.3/1.7) with tokens as the cipher units (U3), capping tokens-per-letter at a large number. The expected type count under a Trithemian scheme is a few hundred to a few thousand, not 5,105 with 70% singletons, so the hypothesis is probably falsifiable by vocabulary growth curves alone (compare the type-token curve of Voynich with that of a Polygraphia III encoding of Latin of the same length). ### 3.3 Pelling's verbose-cipher model and the Averlino hypothesis - **Source.** Nick Pelling, "Voynich Averlino hypothesis summary," ciphermysteries.com, 10 Jan 2010, https://ciphermysteries.com/2010/01/10/voynich-averlino-hypothesis-summary — VERIFIED-FETCHED. *The Curse of the Voynich* (2006 book) — UNVERIFIED as a citation (not retrieved). - **Model as fetched.** The word-initial "4o" (EVA qo) pair is linked to North Italian ciphers of 1424-1456 (Modena 1435: 4o = "qua"; Sforza 1450: 4o = "s"), read as a verbose pair within a more sophisticated system. The fetched post omits the full cipher mechanism ("ran out of room"); components such as aiin as numbers or gallows as abbreviation marks are not in the fetched text. Authorship attributed to Antonio Averlino (Filarete), dating 1451-1500, which the post concedes conflicts with the radiocarbon dating. - **Tried computationally?** No solver run located. - **Implementation.** The testable core is "verbose pairs": a fixed set of digraphs (qo, ol, or, ar, al, ee, ch, sh, aiin family) each standing for one plaintext letter. That is exactly the segmentation-plus-substitution problem of 1.9 (Aldarrab & May 2022) and ranked item 3 below. Under this model the effective alphabet after segmentation should be about 20-25 units with Latin-like conditional entropy; if segmentation into verbose units does not raise conditional entropy toward 3.5 bits, the model fails. ### 3.4 The Naibbe cipher (Greshko 2025) - **Source.** Michael A. Greshko, Cryptologia, published online 26 Nov 2025, DOI 10.1080/01611194.2025.2566408, https://tandfonline.com/doi/full/10.1080/01611194.2025.2566408 — VERIFIED (DOI URL seen; the publisher page could not be fetched; exact title not captured). Details from the Sci.News summary (https://www.sci.news/othersciences/linguistics/voynich-manuscript-cipher-14466.html, VERIFIED-FETCHED) and the Voynich Ninja discussion thread 5081 (VERIFIED-FETCHED). - **Mechanism as reported.** A verbose homophonic substitution of Latin or Italian plaintext, mostly as unigrams and bigrams. Six tables weighted roughly 5:2:2:2:1:1; playing cards (78-card tarocchi or 52-card deck variant) select the table; dice or a respacing scheme handle plaintext respacing; tables contain Voynichese word or affix strings built from prefix and suffix inventories following a Zattera-style slot grammar; when a bigram's output collides with a unigram word type the scribe re-encrypts. Designed to be executable with early-15th-century Alpine materials. - **Statistics reported as reproduced.** Character entropy and conditional entropy, glyph and glyph-pair frequencies, token and type lengths, a quasi-Zipfian rank-frequency curve (R^2 about 0.9 for word frequencies, but the tables were fitted to it), slot-grammar word structure, "skewed pairs." Word-type overlap: about 29% of Naibbe types occur in the manuscript and about 45% of manuscript types can be generated. Reproduces Currier B only incompletely. - **Reported failures and critiques (thread).** Does not reproduce long-range correlations (random-walk analysis); has no line-position awareness so cannot reproduce line-initial/line-final effects; the primary-table dominance means the most frequent tokens leak plaintext letters; the fitted-vs-emergent distinction is contested as circular; far more complex than any attested 15th-century cipher. Rozanova & Temerev 2026 (VERIFIED-FETCHED) independently report that a "published Voynich-imitating cipher" matches the low entropy, unit scale and weak token order but not the edge-glyph coupling or the 70% singleton rate. - **Why it matters.** It is the first published cipher that reproduces the low-entropy, slot-structured profile from a Latin plaintext. It is therefore (a) the right C4 control and (b) the right target for a solver: if a Naibbe-family solver cannot recover Latin from Naibbe-enciphered Latin at 37,000 tokens, no inference about Voynich is possible; if it can, running it on Voynich is a real test. - **Implementation.** Re-implement the generator from the paper (needs the paper itself; the thread gives the architecture). Solver side: a verbose homophonic solver where each cipher unit is a whole token or an affix, with the 5:2:2:2:1:1 prior as a sparsity prior on table usage. Use the leakage noted in the thread: under the primary-table dominance, the most frequent Voynich tokens should align to the most frequent Latin letters/bigrams; the Zipf-fitted tables make that alignment testable by rank correlation with Latin letter frequencies. ### 3.5 Rugg's Cardan grille (2004) and Timm's self-citation generator (2014-2026) - **Sources.** Gordon Rugg, "An Elegant Hoax? A Possible Solution to the Voynich Manuscript," Cryptologia 28(1), January 2004 — VERIFIED-FETCHED (Stolfi bibliography entry; pages not given). Torsten Timm, "How the Voynich Manuscript was created," arXiv:1407.6639 (2014) — VERIFIED-URL. Timm, "Co-Occurrence Patterns in the Voynich Manuscript," arXiv:1601.07435 (c. 2016) — VERIFIED-FETCHED. Timm, "A possible generating algorithm of the Voynich manuscript," Cryptologia 2020 — VERIFIED-FETCHED (academia.edu listing; the listing shows no co-author, the usual co-author Andreas Schinner is UNVERIFIED here, DOI not captured). Timm, "The Challenge of Analyzing a Dynamic Text: Why the Voynich Manuscript Resists Conventional Interpretation," Cryptologia 2026 — VERIFIED-FETCHED (listing). Klaus Schmeh's two-part critique of Rugg, https://scienceblogs.de/klausis-krypto-kolumne/2016/10/28/has-gordon-rugg-solved-the-mystery-of-the-voynich-manuscript-text-part-2 — VERIFIED-URL. - **Status on Voynich.** Both are generators, not ciphers. Parisel 2026 (VERIFIED-FETCHED) tested a Cardan-grille implementation of Rugg's hypothesis and a parametric slot generator against a four-signature criterion and found neither reproduced all four signatures. Rozanova & Temerev 2026 found a self-citation generator failed on edge-glyph coupling and singleton rate. - **Use here.** Controls C3 (self-citation) and, optionally, a grille generator. Any solver that "finds Latin" in C3 output is overfitting. ### 3.6 Latin abbreviation / shorthand (Tironian) proposals - **Status.** The material contains only forum threads (Voynich Ninja tids 2847, 4008, 5553, 5585; VERIFIED-URL as pages, content not fetched), a Notre Dame manuscript-studies tag page (https://sites.nd.edu/manuscript-studies/tag/tironian/, VERIFIED-URL), and media coverage of Cheshire's and Bax's claims. No peer-reviewed evaluation of an abbreviation model was located: UNVERIFIED. Bowern & Gaskell 2022 is the nearest quantitative result (prefix replacement and vowel removal, both abbreviation-like, are among the closest matches to Voynich word metrics). - **Implementation.** Abbreviation is a many-to-one, lossy encoding and so is better modelled as a scorer change than a key search: train the Latin LM on unexpanded abbreviated Latin (transcriptions that keep sigla and suspensions) and on vowel-stripped Latin, then rerun 1.3/1.7. If Voynich scores rise under the abbreviated LM relative to the expanded one, while C5 does not, that is evidence for abbreviation in the plaintext. --- ## 4. Statistical tests of cipher type ### 4.1 Index of coincidence, kappa, and their homophonic variants - **Sources.** Standard; the Zodiac-340 paper (VERIFIED-FETCHED) lists IC and unigram distribution among its diagnostics. No Voynich-specific IC study was retrieved in this material; IC of Voynichese is widely reported in Zandbergen's analysis pages (https://voynich.nu/analysis.html, VERIFIED-URL, not fetched). - **Status.** Applied to Voynich historically (Friedman/Currier era, via Zandbergen's summary). Under-tried: IC computed on U2 units and on tokens, per section, per line position, and per scribe, with periodic IC (kappa at lags 1..200) to detect any alphabet cycling. - **Implementation.** Trivial. The informative version is the kappa test between lines (same-position glyph coincidence across consecutive lines), which would detect line-keyed polyalphabetic behaviour or copy-from-above behaviour (Timm's clustering). ### 4.2 Periodicity and transposition diagnostics (Zodiac-340 battery) - **Source.** Oranchak, Blake & Van Eycke 2024 — VERIFIED-FETCHED (details in 1.3). - **Tests.** (i) Repeating-bigram count at each period p versus a shuffle distribution; (ii) homophone cycling score (do symbols assigned to the same letter alternate regularly?); (iii) minimum spacing between repeated symbols versus expectation; (iv) pivot/isomorph shapes; (v) a trained cipher-type classifier. - **Tried on Voynich?** None of (i)-(iv) located for Voynich. - **Implementation.** Compute (i) on U1 and U2 streams per paragraph (period 1..line length) and on the within-line position grid of each paragraph (to catch columnar reading). Compute (iii) as the distribution of gaps between repeats of the same unit, compared to C1/C2. Given Timm's finding that similar words recur within one or two lines, expect strong signal at periods equal to line length; the test is whether anything remains after conditioning on that. ### 4.3 Isomorphs - **Sources.** Klaus Pommerening, "Cryptanalysis of the Enigma Without Plugboard," https://www.staff.uni-mainz.de/pommeren/Cryptology/Classic/6_Enigma/EnigAna1.html and the rotor-analysis notes at .../5_Rotor/AnalRot.html — VERIFIED-URL. The Zodiac-340 paper's "pivot" patterns are an isomorph-type diagnostic. - **Assumes.** Two stretches of ciphertext with the same repetition pattern (e.g. ABCA ... DEFD) indicate the same plaintext under different alphabets or keys, or the same plaintext under a polyalphabetic cipher at the same key offset. - **Tried on Voynich?** Not located as a formal study. The abundance of near-repeated tokens (Timm) means isomorph counts will be high; the question is whether isomorphs occur *across* different glyph sets at the same line offsets. - **Implementation.** Enumerate repetition patterns (pattern-word signatures) for all windows of length 6-12 units; count pattern pairs that share a signature but no glyphs; compare with C1-C5. A significant excess of glyph-disjoint isomorphs at matched line positions would point to a position-keyed substitution, which no existing model proposes. Effort is low (section 6, item 8). ### 4.4 Unicity distance for Voynich-sized text - **Sources.** Joachim von zur Gathen, "Unicity Distance of the Zodiac-340 Cipher," HistoCrypt proceedings, https://ecp.ep.liu.se/index.php/histocrypt/article/download/395/353 — VERIFIED-FETCHED. Method: key entropy summed over components (homophonic key ~296 bits; sectioning ~15; transpositions ~8; up to 25 arbitrary replacements ~244; total ~563 bits) divided by per-symbol redundancy from n-gram entropies of a Zodiac-specific corpus (1.8 bits/char), giving ~152; robust to entropy choices (132-183). Voynich Ninja thread 4715 (May 2025, VERIFIED-FETCHED): for Brumbaugh's two-stage cipher, a key entropy of ~70 bits and redundancy ~0.5 bit/char gives ~140 characters; participants note unicity is a theoretical bound and the unknown units and language make it hard to apply. - **Tried on Voynich?** Only the forum calculation above. No published unicity budget for the Naibbe, Trithemian or verbose-pair families. - **Worked illustration (my calculation, not from a source).** A Naibbe-like key: 6 tables x ~23 letters, each cell a string drawn from ~500 affix strings, H(K) about 6 x 23 x log2(500) = 1,240 bits; adding ~23^2 bigram cells per table at the same cost gives ~30,000 bits. Latin redundancy with a 23-letter alphabet and a 5-gram entropy near 2.2 bits/char is ~2.3 bits/char. Unicity about 30,000 / 2.3 = 13,000 plaintext characters, well under the ~150,000+ characters the manuscript would carry at two glyphs per letter. So information-theoretically such a cipher is recoverable from the manuscript; the obstacle is search, not key size. Nulls and free respacing add per-use entropy and can push unicity up proportionally; this is the one part of the budget that can make a verbose cipher genuinely unrecoverable, and it should be computed for each proposed family. - **Implementation.** A short script: enumerate key components per family, sum log2 counts, divide by redundancy measured from the period LM. Report alongside every solver run. ### 4.5 Machine-learning cipher-type classification - **Sources.** Ernst Leierzopf, Nils Kopal, Bernhard Esslinger, Harald Lampesberger, Eckehard Hermann, "A Massive Machine-Learning Approach for Classical Cipher Type Detection Using Feature Engineering," HistoCrypt 2021, https://ecp.ep.liu.se/index.php/histocrypt/article/view/164 — VERIFIED-URL. "Detection of Classical Cipher Types with Feature-Learning Approaches," AusDM 2021 preprint, https://www.cryptool.org/download/ncid/Detect-Classical-Cipher-Types-with-Feature-Learning_AusDM2021_PrePrint.pdf — VERIFIED-URL. Neural Cipher Identifier (NCID) online tool, https://legacy.cryptool.org/en/cto/ncid — VERIFIED-URL. Leierzopf's MA thesis, https://www.cryptool.org/media/publications/theses/MA_Leierzopf.pdf — VERIFIED-URL. "Classifying World War II Era Ciphers with Machine Learning," arXiv:2307.00501 — VERIFIED-URL. The Zodiac-340 solution also used a deep-learning classifier that flagged transposition (VERIFIED-FETCHED). - **Assumes.** The ciphertext belongs to one of the trained classes (ACA cipher types in NCID); features such as IC, periodic IC, digraph statistics, repeat patterns, and letter-count statistics. - **Tried on Voynich?** Not located. NCID's classes are ACA ciphers over the Latin alphabet, so the Voynich stream would need to be mapped to 26 symbols, which is itself a unit decision. - **Implementation.** Do not use NCID directly. Instead train the same feature set (plus conditional entropy, slot-grammar coverage, singleton rate, edge-MI) on synthetic ciphers generated from Latin/Italian: monoalphabetic, Tranchedino homophonic, nomenclator-heavy, null-laden, verbose-pair, Naibbe, Trithemian, abbreviation-only, columnar transposition, and the two gibberish generators. A gradient-boosted classifier on 200-token windows gives a posterior over families for each Voynich page. This is the cheapest way to turn the many published "Voynich is close to X" statistics into a single comparable statement. ### 4.6 Bowern-lab metric batteries and Zandbergen's tests - **Sources.** Bowern & Gaskell 2022 (VERIFIED-FETCHED; 42 metrics, 22 manipulations, Euclidean distance on z-scores). Lindemann & Bowern, "Character Entropy in Modern and Historical Texts: Comparison Metrics for an Undeciphered Manuscript," arXiv:2010.14697 (2020) — VERIFIED-URL. Gaskell & Bowern, "Gibberish after all? Voynichese is statistically similar to human-produced samples of meaningless text," VOY2022, https://ceur-ws.org/Vol-3313/paper4.pdf — VERIFIED-URL; code at https://raw.githubusercontent.com/danielgaskell/voynich/main/README.md — VERIFIED-URL. Alexander Boxer, "Fingerprinting Gibberish: A Quantitative Comparison of the Voynich and Sloane MS 3188," VOY2022 — VERIFIED (programme). Zandbergen's analysis pages, https://voynich.nu/analysis.html and https://voynich.nu/a3_para.html — VERIFIED-URL, not fetched. James Tauber, "What Counts as a Character? Alphabet Hypotheses and the Compression Baseline," https://voynich.jtauber.com/posts/alphabet-hypotheses/ — VERIFIED-URL, not fetched. - **Status.** Applied to Voynich. The gap is that these batteries compare Voynich with *manipulated plaintext*, not with *solver behaviour*. The Gaskell repository is a ready-made source of gibberish controls. --- ## 5. Voynich-specific work since 2015 (summary of what has been done) - Bowern & Lindemann, "The Linguistics of the Voynich Manuscript," Annual Review of Linguistics 7 (2021) 285-308, DOI 10.1146/annurev-linguistics-011619-030613, preprint https://ling.auf.net/lingbuzz/005415 — VERIFIED-FETCHED (via the lab page). Review of character- and word-level statistics; concludes the text is more language-like than random but with anomalies. - Sterneck, Polish & Bowern, "Topic Modeling in the Voynich Manuscript," arXiv:2107.02858 (2021) — VERIFIED-FETCHED. NMF topics track scribe and section; stable at 20-word samples; read as evidence of meaningful content. - Bowern & Gaskell 2022; Gaskell & Bowern 2022 — see 4.6. - Zattera 2022 (slot alphabet and grammar; multiple "languages") — see section 0. - Patrick Feaster, "Rightward and Downward Grapheme Distributions in the Voynich Manuscript," VOY2022 — VERIFIED (programme). Caruana, Layfield & Abela, "An Analysis of the Relationship between Words within the Voynich Manuscript," VOY2022 — VERIFIED (programme). Luke Lindemann, "Crux of the MATTR: Voynichese Morphological Complexity," VOY2022 — VERIFIED (programme). Farrugia, Layfield & van der Plas on scribe identification — VERIFIED (programme). Schmeh & Dunin, "The Voynich Manuscript Compared with Other Encrypted Books" (VOY2022 and the March 2026 Zoom conference, VERIFIED-FETCHED agenda; ~120 encrypted books compared). - Layfield et al., conference report on VOY2022, HistoCrypt proceedings, https://ecp.ep.liu.se/index.php/histocrypt/article/download/700/606 — VERIFIED-FETCHED. 16 papers from 32 submissions. - Timm: 2014, 2016, 2020, 2021 review, 2024 "discussion of text creation hypotheses" (Cryptologia), 2026 "dynamic text" (Cryptologia), plus critiques of the five-scribe hypothesis and of Layfield & Davis — VERIFIED-FETCHED (academia listing). - Montemurro & Zanette, "Keywords and Co-Occurrence Patterns in the Voynich Manuscript: An Information-Theoretic Analysis," PLoS ONE 2013, https://pmc.ncbi.nlm.nih.gov/articles/PMC3689824 — VERIFIED-URL. - Zandbergen's STA (Super Transliteration Alphabet), https://voynich.nu/extra/sta.html — VERIFIED-FETCHED: a synthetic superset alphabet (ligatures as single units, family letter + member digit), built on extended EVA and v101; 235 EVA-based characters and combinations found in the ZL file; convertible to nearest basic EVA with a bitrans rules file; STA files exist for CD, FG, IT, VT, ZL, GC and RF transliterations. The analytical "Super Alphabet" is at https://voynich.nu/extra/sta-aaa.html (VERIFIED-URL). This is the right basis for unit hypothesis U2. - Rozanova & Temerev, arXiv:2608.17096 (Aug 2026) — VERIFIED-FETCHED; see section 0. - Parisel, arXiv:2604.19762 (2026) — VERIFIED-FETCHED; see section 0. - Turenne, arXiv:2609.20835 (2026) — VERIFIED-FETCHED; position-dependent probabilistic grammars, phonetic comparison (closer to Hebrew/Arabic than Indo-European), LLM image-text alignment with Pseudo-Apuleius herbals; concludes "generative imitation." - Layfield & Davis, Digital Medievalist (2026), https://journals.openedition.org/digitalmedievalist/2331 — VERIFIED-URL via IFLScience (VERIFIED-FETCHED): LSA used to argue the manuscript was originally loose bifolia later bound out of order; contested by Timm. - Self-published Zenodo items (Bezvesilnyi "Voynich Protocol v3.0," DOI 10.5281/zenodo.18518992; Piesiak "DECRYPTED ... 99.8% Confidence") — VERIFIED-FETCHED as existing; not peer-reviewed; the Bezvesilnyi corpus count of 116,755 words is inconsistent with any transliteration and the Piesiak claims give no reproducible procedure. Listed only so they are not mistaken for prior art. - Nick Pelling's blog (ciphermysteries.com) — VERIFIED-FETCHED for the two posts above. Emma May Smith (agnosticvoynich) and Elmar Vogt (voynichthoughts) — UNVERIFIED: the one search that targeted them returned only forum threads and unrelated arXiv items; no page was fetched. - Voynich Conference 2026 (University of Malta, online, 9 Dec 2026; solutions not accepted) — VERIFIED-FETCHED. **Net assessment.** Since 2015 the Voynich-specific literature has produced strong *descriptive* constraints (entropy, slot grammar, singleton rate, line effects, edge coupling, directional asymmetry) and two generator classes that partly reproduce them (self-citation; Naibbe). It has produced almost no *solver* work: no published run of a modern homophonic/verbose solver with controls, no transposition diagnostics, no cipher-family classifier, no unicity budgets, and no application of the Snyder/Luo lost-language models. That is the gap the ranked list addresses. --- ## 6. Ranked top-10 untried or under-tried techniques Effort is a rough estimate for a working Python implementation (numpy/numba) on a 2-core machine, excluding corpus collection unless stated. Each item assumes the control set C1-C6 from section 0 already exists (building the controls is itself ~15-25 h and is the prerequisite for everything below). **1. Verbose homophonic solver run with a composite n-gram + entropy score and period-specific LMs, with full controls.** This is the one experiment that the entire Zodiac/DECRYPT/Lasry tool-chain was built for and that nobody has published on Voynich. Units: U2 slot units and, in a second pass, Zattera-style affix strings; each unit maps to a Latin letter, a letter pair, or null, with a cap on units per letter (Lasry) and an entropy term (AZdecrypt) so the solver cannot collapse the alphabet. Scorer: 5- and 6-gram LMs from 14th-15th-century Latin and Italian, plus an abbreviated-Latin LM (Megyesi 2023 shows period matters; Lasry's French letter only broke with a period corpus and 5-grams). Run per Currier language and per Zattera section. Decision rule: the Voynich score must exceed the C3 and C4 scores at the same length, and the solver must recover C5 at that length. Even a negative result is publishable because it bounds which key families are searchable at this corpus size. Effort: 30-40 h (solver 15, LMs 8, runs and reporting 10). **2. Transposition and periodicity diagnostics (the Zodiac-340 battery).** Repeating-bigram counts by period, repeat-gap distributions, homophone-cycling scores, pivot/isomorph shapes, all computed on U1 and U2 streams per paragraph and on the line-position grid, against C1-C5 distributions. This has never been run on Voynich and it directly tests a family (substitution composed with a regular transposition or reading order) that no Voynich model currently includes, while also quantifying Timm's within-line repetition as a periodic signal that any later solver must condition on. Effort: 6-10 h. **3. Joint segmentation-and-substitution ("verbose cipher") inversion.** Pelling's verbose pairs, Naibbe's affix tables and Bowern-Gaskell's bigraph encodings all say the same thing: the cipher unit is a variable-length glyph string. Aldarrab & May 2022 show how to learn segmentation and substitution jointly for numerical ciphers; the same EM/beam formulation applies to glyph strings with a lexicon of candidate units (every 1-3-unit substring with frequency above a threshold). The target statistic is simple: after the learned segmentation, conditional entropy of the segmented stream should approach 3.5 bits if Latin lies underneath, and the segmented alphabet size should be 20-30. If no segmentation moves entropy materially, verbose substitution is ruled out for the tested unit lexicon. Effort: 25-40 h. **4. Alphabet-unit hypothesis selection by solver performance and compression.** The Copiale lesson: decide which graphemes carry information before solving. Candidates: basic EVA, STA families, Zattera slots, Rozanova-Temerev learned units, and null-class partitions (gallows as nulls, line-initial glyphs as nulls, i/n tails as nulls). For each: compression baseline (Tauber), unicity budget (item 7), and the score of item 1's solver on Voynich versus C5 under the same partition. The output is a short table that every subsequent experiment can cite for its unit choice. Effort: 10-15 h given item 1. **5. Cipher-family classification trained on historically attested synthetic ciphers.** Leierzopf/Kopal-style feature engineering, but trained on Latin and Italian enciphered with Tranchedino keys from DECODE, nomenclator-heavy keys, null-laden keys, verbose pairs, Naibbe, Trithemian word-for-letter, abbreviation-only, columnar transposition, and the two gibberish generators, on 200-token windows. Features: IC and periodic IC, conditional entropy, slot-grammar coverage, singleton rate, edge mutual information, repeat-gap statistics, type-token curve slope. A gradient-boosted classifier yields a per-page posterior over families and, importantly, a calibrated "none of the above" rate from held-out families. This converts a decade of "Voynich resembles X" results into one comparable number per hypothesis. Effort: 15-25 h. **6. Period- and register-specific language models for scoring, including abbreviated Latin.** Cheap to add to item 1 but worth listing separately because Megyesi 2023 and Lasry's letter both show that the LM, not the search, was the deciding factor. Build character LMs from (a) 14th-15th-century Latin herbal and medical texts, (b) Italian and Occitan of the same period, (c) unexpanded abbreviated Latin, (d) vowel-stripped variants. Measure how much each LM changes the solver's Voynich-versus-control margin. Corpus collection is the real cost. Effort: 10-15 h plus corpus gathering. **7. Unicity-distance budgets for each candidate cipher family at Voynich length.** Von zur Gathen's method applied to the Naibbe, verbose-pair, Trithemian, Tranchedino-plus-nulls and abbreviation families, with key entropy summed per component and redundancy from the period LMs, and with the per-use entropy of nulls and free respacing modelled explicitly. This tells the program which families are recoverable in principle from ~37,000 tokens and which are not, before any solver time is spent; the forum calculation for Brumbaugh's cipher is the only prior attempt. Effort: 4-8 h. **8. Isomorph and glyph-disjoint repetition-pattern analysis across lines.** Enumerate repetition signatures of 6-12-unit windows and count signature matches that share no glyphs, by line offset and by position-in-line, against C1-C5. An excess of glyph-disjoint isomorphs at matched positions would indicate a position-keyed or line-keyed substitution, a family nobody has proposed for Voynich because nobody has looked; a deficit relative to self-citation output would be evidence for copying rather than keying. Effort: 6-10 h. **9. Luo et al. 2021-style undersegmented decipherment with phonetic priors, run as language identification.** The method was built for exactly the Voynich situation (unknown script, uncertain segmentation, candidate language set) and has not been applied to it. Run on U1 and U2 with spaces removed and with spaces as soft boundaries, against Latin, Italian, Occitan, German, Czech, Greek, Hebrew and Arabic lexica, and against C3/C4 where the correct answer is "no language." Report the ranked list with the control-calibrated confidence, and repeat on a vowel-stripped Latin control to see how the phonetic prior behaves when the script is an abjad or an abbreviation system. This also answers Zandbergen's demand that Hauer & Kondrak's Hebrew result be reproduced on other transliterations and sections with a stronger method. Effort: 30-50 h (re-implementation from the paper; no GPU needed at this scale but training is slow on two cores). **10. Constrained word-level (nomenclator) alignment with split-sample validation.** Tranchedino keys carried 80-plus word entries and the Naibbe tables carry whole tokens, so some fraction of Voynich tokens may be code words rather than enciphered letters. A Snyder-style Bayesian alignment between the most frequent Voynich tokens (top 200-500) and a Latin herbal/medical lexicon, constrained by token position (label versus paragraph, line-initial versus medial), by page section (Zattera's five "languages"), and by a min-cost-flow one-to-one constraint (Luo 2019), fitted on Currier A and tested on Currier B and on the labels. The self-published "decryptions" in section 5 do a degenerate version of this with no held-out test; the rigorous version has not been done. Expected outcome is probably negative, but it is the only direct test of the nomenclator component that every attested 15th-century cipher contained. Effort: 20-30 h. **Not ranked but worth noting.** Neural seq2seq family detectors (section 1.9) are a heavier way to do item 5 and would be the next step if item 5 gives a non-trivial posterior on a specific family. Topic modelling, character-entropy comparison, monoalphabetic-plus-anagram solving, and word-metric batteries have all been done and should be reused as controls rather than repeated.