# Methods survey: sequence-analysis techniques from genomics, signal processing, complex systems, animal communication and epigraphic computing, assessed for the Voynich text

Prepared 2026-10-22 from the collected search/fetch record (a7458c0c48153fa2b.full.md) plus six verification searches. Every citation is tagged **VERIFIED** (title and URL/DOI seen in a fetched page or search result) or **UNVERIFIED** (cited from memory; details may be wrong). Where a URL was seen but volume/page/DOI details were not, the tag says so. Nothing below reports a result that was not in the collected material or the cited source; where I summarise a paper's findings from memory rather than from a fetched page, the sentence says "from memory".

Techniques the research program has **already run** (mutual-information decay, gzip compressibility, BPE merge curves, repeated-phrase counts, 2/3-state HMM, form–function edit-distance-vs-context tests) are described for completeness but are not ranked as untried.

---

## 0. Working conventions used in every implementation note

**Corpus.** An IVTFF-format EVA transliteration (Zandbergen's ZL or Takahashi's TT, both distributed via voynich.nu). Strip inline comments, drop tokens containing uncertain readings (`?`, `[a:b]` alternatives) unless the test needs them; Layfield et al. 2020 kept 32,216 high-certainty word tokens this way (VERIFIED, https://preview.aclanthology.org/fix_video/2020.lt4hala-1.11.pdf). Keep line, paragraph, folio, Currier-language (A/B), section (herbal, astronomical, biological, pharmaceutical, stars) and, where available, Davis's scribe assignments as metadata; run everything per section as well as globally, since A and B differ on most statistics (Bowern & Gaskell 2022 found Currier B closer to their cipher controls than A; VERIFIED, https://ceur-ws.org/Vol-3313/paper6.pdf).

**Units.** Three alphabets, always reported side by side, because the EVA encoding itself drives character-level results (forum critiques of quimqu's n-gram study made exactly this point; VERIFIED, https://www.voynich.ninja/thread-4763.html):
- U1: raw EVA characters;
- U2: glyph-merged alphabet (`ch`, `sh`, `cth/ckh/cph/cfh`, `ii/iii`, `in/iin/iir` etc. as single units);
- U3: word tokens (space-delimited; keep a flag for uncertain spaces).

**Null models (shuffle hierarchy).** N0 global character shuffle; N1 character shuffle within word; N2 global word shuffle; N3 word shuffle within line; N4 line shuffle within page; N5 page shuffle; N6 text generated by the Timm–Schinner self-citation procedure (copy-and-modify of earlier words); N7 the Zipf-preserving, long-range-correlated fractional-Gaussian-noise surrogate of Montemurro & Degli Esposti 2026 (VERIFIED, arXiv:2603.02213, https://arxiv.org/abs/2603.02213), which keeps word frequencies and the DFA exponent and randomises short-range structure. The point of the hierarchy is to locate the scale at which a statistic's signal lives.

**Controls.** Positive: 37k-token excerpts of medieval Latin, Italian and German (and one non-Indo-European text), passed through identical preprocessing. Cipher controls: the same texts after simple substitution, vowel deletion, phoneme-class merging, bigraph encoding and null insertion, i.e. the Bowern & Gaskell 2022 manipulation set (VERIFIED). Negative: N6 output and, if obtainable, the human-written gibberish samples of Gaskell & Bowern 2022 (VERIFIED, https://ceur-ws.org/Vol-3313/paper4.pdf; code/data repository https://raw.githubusercontent.com/danielgaskell/voynich/main/README.md). Every test below reports (Voynich A, Voynich B) against (language controls, cipher controls, gibberish controls) under the same statistic; a result is "defensible" only if Voynich separates from at least one control family by more than the shuffle-derived confidence interval.

**Compute.** All estimates assume Python/NumPy, optional numba, 2 cores, no GPU. A 37k-token / ~190k-character sequence fits comfortably in memory; the costly items are all-pairs comparisons and MCMC.

---

## 1. Sequence analysis

### 1.1 k-mer spectra and Zipf / Zipf–Mandelbrot fits (the DNA-linguistics controversy)
- **Key citations.** Mantegna, Buldyrev, Goldberger, Havlin, Peng, Simons, Stanley, "Linguistic features of noncoding DNA sequences", Phys. Rev. Lett. 73, 3169 (1994) — UNVERIFIED (details from memory; only the comments were seen). Bonhoeffer, Herz, Boerlijst, Nee, Nowak, May, "Comment on 'Linguistic features of noncoding DNA sequences'" — VERIFIED as arXiv:adap-org/9503002 (https://arxiv.org/pdf/adap-org/9503002); authorship from memory. Niyogi & Berwick, "A note on Zipf's law, natural languages, and noncoding DNA regions", MIT AI memo (1995) — VERIFIED as arXiv:cmp-lg/9503012 (https://arxiv.org/pdf/cmp-lg/9503012) and MIT DSpace 1721.1/6634; authorship from memory.
- **Question answered.** Does the rank–frequency spectrum of fixed-length "words" (k-mers) look Zipfian, and does Zipfian behaviour discriminate language from non-language?
- **Prior application.** Zipf fits on Voynich word tokens are routine (Gaskell & Bowern 2022 include a Zipf fit among their metrics, VERIFIED; Turenne 2026 reports Zipf-like fits, VERIFIED, https://arxiv.org/abs/2609.20835). The lesson of the DNA controversy is the relevant part: the comments showed that Zipf-like spectra arise from many non-linguistic processes (random texts with unequal letter probabilities, Markov sources), so a Zipf fit is not evidence for language. The same argument, made by Sproat for ancient symbol systems (section 5.1), applies to Voynich.
- **Implementation (37k tokens).** Compute rank–frequency for U3 tokens and for U2 k-mers, k = 2..6, ignoring spaces; fit Zipf–Mandelbrot f(r) = C/(r+b)^a by maximum likelihood with a discrete power-law estimator and a Kolmogorov–Smirnov goodness-of-fit bootstrap (not least squares on log–log). Nulls: N0 and N1 produce "random text" spectra; N6 produces copy-and-modify spectra. The only informative output is the exponent and curvature *relative to controls at equal token count*, since Heaps/Zipf parameters depend on sample size. Treat this as a calibration step for other tests, not a finding. ~3 h.

### 1.2 Mutual-information decay / long-range correlations — ALREADY RUN
- **Key citations.** Li, "Mutual information functions versus correlation functions", J. Stat. Phys. 60, 823 (1990) — UNVERIFIED. Li & Kaneko, "Long-range correlation and partial 1/f^α spectrum in a noncoding DNA sequence", Europhys. Lett. 17, 655 (1992) — UNVERIFIED (search returned only secondary Li material: https://repository.cshl.edu/id/eprint/31490/). Peng, Buldyrev, Goldberger, Havlin, Sciortino, Simons, Stanley, "Long-range correlations in nucleotide sequences", Nature 356, 168 (1992) — VERIFIED title (https://cris.biu.ac.il/en/publications/long-range-correlations-in-nucleotide-sequences/), volume/page from memory. Montemurro & Zanette, "Keywords and co-occurrence patterns in the Voynich manuscript: an information-theoretic analysis", PLoS ONE 8(6): e66344 (2013) — VERIFIED (https://pmc.ncbi.nlm.nih.gov/articles/PMC3689824), DOI from memory 10.1371/journal.pone.0066344.
- **Prior application.** Montemurro & Zanette 2013 computed word-level long-range correlations and an entropy-based keyword measure on Voynich and found structure comparable to natural-language texts (summary from memory; the fetched material gives only the title). The program has run MI decay already.
- **Implementation note.** For any re-run: compute I(d) between symbols at lag d on U2 with Miller–Madow or NSB bias correction, and subtract the mean I(d) of 200 N0 shuffles rather than reading raw I(d), since finite-sample bias at 190k symbols is of the same order as the language signal beyond d ≈ 50. Compare the decay form (power law vs exponential) against N7, which by construction preserves long-range scaling without short-range grammar.

### 1.3 Detrended fluctuation analysis (DFA)
- **Key citation.** Peng, Buldyrev, Havlin, Simons, Stanley, Goldberger, "Mosaic organization of DNA nucleotides", Phys. Rev. E 49, 1685 (1994) — VERIFIED title (https://cris.biu.ac.il/en/publications/mosaic-organization-of-dna-nucleotides/), volume/page from memory.
- **Question answered.** Does a scalar series derived from the symbol stream show scale-free fluctuations (exponent α ≠ 0.5), i.e. long memory, after removing local trends?
- **Prior application.** The Voynich-specific search for DFA/Hurst returned no DFA study on Voynich; DFA has been applied to sentence-length and sentiment series in literary texts (VERIFIED as existing, e.g. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5029871/). Montemurro & Degli Esposti 2026 use the DFA exponent as the quantity their surrogate preserves (VERIFIED). Landini's 2001 Hurst/spectral study is the closest prior work (1.4). **Not applied in the modern DFA form.**
- **Implementation.** Build several series from the EVA stream: word length per token; gallows indicator (0/1 per token); a "DNA-walk" style ±1 encoding for each binary partition of U2 (vowel-like vs consonant-like classes from the 2-state HMM; benched vs plain); line length per line. Compute DFA-1 and DFA-2 fluctuation functions F(n) for n = 4 … N/4 in log-spaced windows, fit α on the straight region, report with 95% CI from 500 N2 (word-shuffled) and N4 (line-shuffled) surrogates. Decisive comparison: α_Voynich vs α of language controls and of N6. Crossover scales (where α changes) are the useful byproduct: a crossover at ~line length or ~paragraph length is layout, not language. ~5 h including surrogates.

### 1.4 Hurst exponent (rescaled range) and spectral analysis
- **Key citation.** Landini, "Evidence of linguistic structure in the Voynich manuscript using spectral analysis", Cryptologia 25(4), 275–295 (2001) — UNVERIFIED (no direct hit in two searches; existence attested indirectly because Caruana et al. 2022 discuss Landini's spectral work, VERIFIED via fetched summary of https://ceur-ws.org/Vol-3313/paper8.pdf).
- **Question answered.** Same as DFA, with the classical R/S estimator and the power spectrum of the symbol series; Landini reportedly found 1/f-like spectra and Hurst exponents consistent with language (from memory).
- **Prior application.** Applied to Voynich (Landini 2001). Under-tried in the sense that it predates surrogate-based significance testing; R/S is biased for short series and sensitive to the arbitrary numeric encoding of symbols.
- **Implementation.** Re-run only as a check on 1.3: Welch periodogram of the same series, fit the 1/f^β slope, and confirm β ≈ 2α − 1 holds, which it will not if the "signal" is an encoding artefact. ~2 h, bundled with 1.3.

### 1.5 Lempel–Ziv complexity
- **Key citation.** Lempel & Ziv, "On the complexity of finite sequences", IEEE Trans. Inf. Theory 22, 75–81 (1976) — UNVERIFIED.
- **Question answered.** How many distinct phrases does the LZ76 parse of the sequence produce, normalised by the value for a random sequence of the same alphabet and length; a proxy for entropy rate that is less compressor-dependent than gzip.
- **Prior application.** Not applied to Voynich as LZ76 complexity specifically; gzip (an LZ77 variant) has been run by the program.
- **Implementation.** Implement the Kaspar–Schuster LZ76 parse (O(N²) worst case but fine for 190k symbols with a suffix-automaton or simply a dictionary of seen substrings). Normalise c(N)·log_k(N)/N. Compute on U1, U2 and on U3 (word IDs as symbols). Nulls N0–N5 locate the scale: if the Voynich value drops sharply only under N3 (within-line word shuffle), the regularity is intra-line. Report alongside 1.6. ~3 h.

### 1.6 Entropy-rate estimators (LZ78 / PPM / context-tree weighting) and Hilberg extrapolation
- **Key citations.** Willems, Shtarkov, Tjalkens, "The context-tree weighting method: basic properties", IEEE Trans. Inf. Theory 41, 653–664 (1995) — UNVERIFIED. Takahira, Tanaka-Ishii, Dębowski, "Entropy rate estimates for natural language — a new extrapolation of compressed large-scale corpora", Entropy 18(10), 364 (2016) — VERIFIED (https://www.mdpi.com/1099-4300/18/10/364/xml); workshop version "Upper bound of entropy rate revisited", W16-4124 — VERIFIED (https://aclanthology.org/W16-4124/).
- **Question answered.** What is the per-symbol entropy rate h, and how does the compressed size scale with length (Hilberg/stretched-exponential ansatz r(n) = A·n^β + h·n)? Natural languages show β ≈ 0.8–0.9 and h well below the first-order entropy.
- **Prior application.** Character-entropy comparisons on Voynich exist (Lindemann & Bowern, "Character entropy in modern and historical texts", arXiv:2010.14697, VERIFIED https://ar5iv.labs.arxiv.org/html/2010.14697, and the Yale PDF), as do n-gram perplexity curves (quimqu 2025, VERIFIED). CTW/PPM estimates with Takahira-style extrapolation: **not applied**.
- **Implementation.** Use PPMd (via `pyppmd` or the `ppmd-cffi` wheel) and a CTW implementation (a few hundred lines; depth 8–16 on U2) on prefixes of length 2^k, k = 10 … 17; fit the Hilberg ansatz by nonlinear least squares; bootstrap over 50 random prefix start points. Compare (h, β) with language controls at equal length, with cipher controls (substitution preserves h; vowel deletion and bigraph encoding lower it, per Bowern & Gaskell 2022), and with N6. The interesting quantity is β, which separates "low entropy because of a small repetitive vocabulary" from "low entropy because of long-range structure". ~6 h.

### 1.7 Normalized compression distance (NCD) for clustering sections, hands and languages
- **Key citation.** Cilibrasi & Vitányi, "Clustering by compression", IEEE Trans. Inf. Theory 51(4), 1523–1545 (2005) — VERIFIED (https://www.cwi.nl/documents/199318/cluster.pdf; arXiv:cs/0312044), DOI 10.1109/TIT.2005.844059 from memory.
- **Question answered.** Which pages/sections are "closest" under a universal similarity measure, with no feature engineering?
- **Prior application.** Search found no published NCD study of Voynich; the Voynich Ninja thread "Inside the Voynich Network: graph analysis" was fetched and does not contain NCD work (VERIFIED, https://www.voynich.ninja/thread-4998-page-4.html). **Not applied** in print. Note that NCD between Voynich and natural languages is uninformative because the alphabets differ; NCD *within* the manuscript is the use case.
- **Implementation.** Compute NCD(x,y) = (C(xy) − min(C(x),C(y))) / max(C(x),C(y)) with a compressor whose window exceeds the inputs (xz/lzma or PPMd, not gzip, which has a 32 kB window) over all folio pairs (≈ 225 folios → 25k pairs, seconds). Cluster with neighbour joining or UPGMA; compare the tree to Currier A/B, to the five illustration sections, to quire boundaries and to Davis's scribal hands (section 7.1) with an adjusted Rand index. Null: the same tree built from N5 page-shuffled text has no structure; the discriminating check is whether the compression tree agrees better with scribe than with section. ~4 h.

### 1.8 "Zipping" language classification
- **Key citation.** Benedetto, Caglioti, Loreto, "Language trees and zipping", Phys. Rev. Lett. 88, 048702 (2002) — VERIFIED as arXiv:cond-mat/0108530 (https://arxiv.org/abs/cond-mat/0108530), journal details from memory. Goodman, "Extended comment on Language trees and zipping" — VERIFIED (https://arxiv.org/pdf/cond-mat/0202383); reply — VERIFIED (https://arxiv.org/abs/cond-mat/0203275).
- **Question answered.** Given a reference corpus, which reference is most "compressible together" with the unknown text? Useful for authorship/hand attribution within a single alphabet.
- **Prior application.** Not applied to Voynich. Goodman's comment is worth reading first: he showed that simple n-gram models outperform zipping at the same task, so this method's only advantage is implementation simplicity.
- **Implementation.** Use it only as an internal attribution tool: take each scribe's pages (Davis 2020) as a reference, compute the cross-entropy increment C(ref+page) − C(ref) for every held-out page, and test whether pages are assigned to their own scribe above chance. The same experiment done with a character 5-gram model (section 6.1) is the stronger version. ~3 h.

### 1.9 Zipf-preserving long-range surrogate as a null model
- **Key citation.** Montemurro & Degli Esposti, "A Zipf-preserving, long-range correlated surrogate for written language and other symbolic sequences", arXiv:2603.02213 (2026) — VERIFIED (fetched).
- **Question answered.** Which Voynich statistics survive when word frequencies and the DFA exponent are held fixed and everything else is randomised? This is a null, not a test, and it is new enough that no Voynich study has used it.
- **Implementation.** Generate fractional Gaussian noise with the Voynich DFA exponent (from 1.3), rank-map it onto the empirical word histogram, produce 200 surrogates. Then re-run every word-level statistic the program already has (MI decay, repeated-phrase counts, burstiness) against this null instead of plain shuffles. A statistic that is *not* reproduced by N7 is evidence of structure beyond "Zipf plus memory", which is exactly the structure Timm–Schinner copy-and-modify would produce; a statistic that *is* reproduced is uninformative. ~6 h including re-runs.

---

## 2. Motif discovery and segmentation

### 2.1 MEME / Gibbs-sampling motif finders
- **Key citations.** Bailey & Elkan, "Fitting a mixture model by expectation maximization to discover motifs in biopolymers", ISMB 1994 — UNVERIFIED. Lawrence, Altschul, Boguski, Liu, Neuwald, Wootton, "Detecting subtle sequence signals: a Gibbs sampling strategy for multiple alignment", Science 262, 208–214 (1993) — UNVERIFIED.
- **Question answered.** Are there position-weight-matrix motifs (fixed-width, degenerate patterns) that are over-represented in a set of sequences, e.g. line-initial, line-final or paragraph-initial contexts?
- **Prior application.** Not applied to Voynich. Positional glyph biases are known from Feaster, "Rightward and downward grapheme distributions in the Voynich manuscript" (VOY2022, VERIFIED listing, https://ceur-ws.org/Vol-3313/) and from Reddy & Knight 2011 (VERIFIED, https://aclanthology.org/W11-1511/), but as per-position frequencies, not as motifs.
- **Implementation.** Treat each line as a "sequence"; run a zero-or-one-occurrence-per-sequence (ZOOPS) PWM finder of width 3–6 on U2 against a background 2nd-order Markov model estimated from the whole text. MEME itself (the `meme` binary) accepts custom alphabets via `-alph`, so no reimplementation is needed; a Gibbs sampler in NumPy is ~150 lines if a Python-only solution is preferred. Score motifs by E-value against the Markov background, then test positional enrichment (line-initial vs mid-line) with a permutation test under N3. Known structure (paragraph-initial gallows, line-final `m`/`g`) is the positive control: if the finder cannot recover those, discard it. New information would be mid-line motifs with no positional explanation. ~6 h.

### 2.2 HMM segmentation (unsupervised Baum–Welch; "gene-finder" style semi-Markov models) — 2/3-STATE VERSION ALREADY RUN
- **Key citations.** Reddy & Knight, "What we know about the Voynich manuscript", ACL LaTeCH 2011 — VERIFIED (https://aclanthology.org/W11-1511/); from memory, they fit a 2-state character HMM that separated glyphs into two classes, and tested word- and page-level statistics. Generalised/semi-Markov HMMs with explicit state-duration distributions are the standard gene-finder architecture (GENSCAN, Burge & Karlin 1997) — UNVERIFIED.
- **Question answered.** Does the stream decompose into a small number of hidden regimes with distinct emission statistics and characteristic lengths?
- **Prior application.** 2-state (Reddy & Knight) and the program's 2/3-state runs. **Under-tried:** more states, state duration modelling, and positional (line-start/line-end) states.
- **Implementation.** Two extensions. (a) Model selection over K = 2 … 12 states on U2 with held-out log-likelihood (train on odd lines, test on even), BIC, and per-section fits; report the state-posterior sequence and the glyph emission profile for each state. (b) A hidden semi-Markov model on U3 where each state emits whole words from a state-specific unigram and has an explicit duration distribution (negative binomial), which is the gene-finder analogue of exon/intron: it tests whether word runs of a given "type" have characteristic lengths. The `hmmlearn` package handles (a); (b) needs a ~300-line forward–backward with duration (or the `hsmmlearn`/`pyhsmm` packages). Nulls: N3 destroys within-line regimes; N6 output trained the same way shows what copy-and-modify produces. ~10 h.

### 2.3 Change-point detection
- **Key citation.** Killick, Fearnhead, Eckley, "Optimal detection of changepoints with a linear computational cost" (PELT), J. Am. Stat. Assoc. 107, 1590–1598 (2012) — UNVERIFIED. The `ruptures` Python package implements PELT, binary segmentation and kernel change-point — UNVERIFIED package reference.
- **Prior application.** A folio-level Bayesian multinomial change-point model in Stan on word-length bins and on 12-/34-topic assignments (hepialidae, Weird Data Science, 2020; VERIFIED, https://www.r-bloggers.com/2020/02/illuminating-the-illuminated-part-four-tempora-mutantur-changepoint-analysis-of-the-voynich-manuscript/), which placed change-points around folios 33 and 38. **Under-tried:** line- or paragraph-granularity, multivariate features, and alignment with scribal hands and quires.
- **Implementation.** Build a per-line feature vector: mean word length, gallows rate, benched-gallows rate, `-edy`/`-aiin` suffix shares, `qo-` prefix share, type/token ratio in a 50-word window, Currier-A/B classifier score. Order lines in manuscript order (and separately in a quire-reordered sequence). Run PELT with an RBF-kernel cost and a penalty selected by the slope heuristic; bootstrap change-point locations with N4 (line shuffle within page) to get a false-alarm rate. Compare detected break positions to (i) illustration-section boundaries, (ii) quire boundaries, (iii) Davis's scribe changes, (iv) folio/page boundaries, using a precision/recall at ±2 lines. The hypothesis worth testing: text-statistical breaks follow scribe/quire rather than illustration section. ~8 h.

### 2.4 Bayesian word segmentation (Dirichlet-process unigram/bigram models)
- **Key citation.** Goldwater, Griffiths, Johnson, "A Bayesian framework for word segmentation: exploring the effects of context", Cognition 112(1), 21–54 (2009) — VERIFIED (https://cocosci.princeton.edu/tom/papers/wordseg3.pdf), DOI 10.1016/j.cognition.2009.03.008 from memory.
- **Question answered.** Given an unsegmented symbol stream, where does an ideal learner with a lexicon prior put word boundaries? Comparing recovered boundaries with the scribe's spaces tests whether Voynich "words" are lexical units at all.
- **Prior application.** Not applied to Voynich in print (the Voynich-specific segmentation search found only a forum post claiming four-slot morphology with "unsupervised boundary tests", unspecified, in the forum's "Slop Bucket" section; VERIFIED existence, https://www.voynich.ninja/archive/index.php/thread-5013.html). Lindemann's MATTR work (VOY2022, VERIFIED listing) addresses morphological complexity with a different tool.
- **Implementation.** Remove spaces from U2 lines; run the DP-unigram and HDP-bigram Gibbs samplers (the `wordseg` package or dpseg, or a ~400-line NumPy implementation) for 2,000 sweeps with annealing; compute boundary precision/recall/F1 against the scribe's spaces, and also token F1. Do the same for the Latin/Italian controls with their spaces removed (expected F1 roughly 0.7–0.85 for the bigram model on natural language, from memory of the literature; measure it rather than assume it). Then interpret: Voynich F1 ≈ control F1 means scribal spaces behave like lexical boundaries; F1 much lower, with systematic over-segmentation at `qo|`, `|aiin`, `ch|ed|y` etc., means the "word" is a composite of smaller recurring units, consistent with a cipher-group or slot-grammar reading; much *higher* F1 than controls means words are more predictable from their edges than real words are, consistent with copy-and-modify generation. Also run with spaces kept but uncertain-space tokens masked, to see where the model wants to split them. ~10 h (sampling dominates; HDP bigram on 190k characters is minutes per sweep in NumPy, so use the compiled `dpseg`/`wordseg` backends).

### 2.5 Adaptor grammars for morphology
- **Key citations.** Johnson, Griffiths, Goldwater, "Adaptor grammars: a framework for specifying compositional nonparametric Bayesian models", NIPS 2007 — UNVERIFIED. Eskander, Rambow, Yang, "Extending the use of adaptor grammars for unsupervised morphological segmentation of unseen languages", COLING 2016 — VERIFIED (https://preview.aclanthology.org/fix_video/C16-1086.pdf).
- **Question answered.** Which prefix–stem–suffix grammar best explains the word list under a nonparametric Bayesian prior, and how large is the induced morpheme inventory?
- **Prior application.** Not applied to Voynich. Stolfi's hand-built "crust–mantle–core" word grammar is the informal predecessor (UNVERIFIED web page).
- **Implementation.** Use the PrefixStemSuffix and the compounding grammars from Eskander et al. with Mark Johnson's `py-cfg` sampler (C++; builds on Linux) on the U2 type list weighted by token frequency; 1,000 sweeps. Outputs: morph inventory size, per-word segmentation, and the sampled grammar's log posterior. Compare the inventory size and the ratio of stems to affixes against Latin/Italian controls segmented the same way, and against N6 text. Measure agreement with Morfessor (2.6) and with BPE merges the program already has; disagreement between the three is itself informative about whether the word structure is positional (slot) or concatenative (morphological). ~8 h.

### 2.6 Morfessor
- **Key citation.** Creutz & Lagus, "Unsupervised models for morpheme segmentation and morphology learning", ACM Trans. Speech Lang. Process. 4(1), art. 3 (2007) — VERIFIED (https://ufal.mff.cuni.cz/~hana/teaching/2011su-morph/creutz-lagus-2007.pdf). Morfessor 2.0 (Virpioja et al. 2013) — UNVERIFIED.
- **Question answered.** MDL-style segmentation of the word list into a lexicon of morphs plus a corpus encoding; the two-part code length is a direct, comparable measure of how "morphological" a word list is.
- **Prior application.** Not applied to Voynich in print.
- **Implementation.** `pip install morfessor`; train Baseline on U2 types with token-count weighting and with the corpus-weight parameter swept over 0.1–10; record lexicon cost, corpus cost, and the number of morphs per word. The comparable quantity across texts is the *cost curve* over the sweep, not a single point. Positive control: Latin should yield a small affix set with high reuse; N6 output should yield many near-duplicate morphs. ~3 h, usually bundled with 2.5.

### 2.7 Byte-pair encoding as an analysis tool — ALREADY RUN
- Merge curves have been computed by the program. One unexploited reading: the rank at which merges stop being "within-word" and start crossing space boundaries, which differs between languages with real word boundaries and symbol systems with arbitrary grouping. No further ranking.

### 2.8 MDL grammars (Sequitur) and grammar induction
- **Key citation.** Nevill-Manning & Witten, "Identifying hierarchical structure in sequences: a linear-time algorithm", J. Artif. Intell. Res. 7, 67–82 (1997) — VERIFIED (https://arxiv.org/abs/cs/9709102; https://jair.org/index.php/jair/article/view/10192). Their follow-up "Compression and explanation using hierarchical grammars" — VERIFIED title only.
- **Question answered.** What is the smallest context-free grammar (one rule per repeated digram, applied recursively) that generates the sequence; how deep is the rule hierarchy; how often are rules reused? Used on DNA and music to expose repeated hierarchical structure.
- **Prior application.** Not applied to Voynich (search found nothing).
- **Implementation.** Sequitur is ~200 lines in Python (several public implementations exist); run on U2 characters with spaces as symbols, and separately on U3 word IDs. Report: grammar size (total symbols in rules), number of rules, maximum and mean rule depth, rule reuse histogram, and the top 50 rules expanded. Compare to the same statistics on controls at equal length and on N3/N6 nulls. The two signals to look for: (i) deep hierarchy with high reuse (language-like), vs (ii) many shallow rules reused locally within a page (copy-and-modify-like, which N6 will show). Also compute "rule span": the mean manuscript distance between uses of each rule; language rules are global, self-citation rules are local. ~8 h including the word-level run and nulls.

---

## 3. Alignment and homology

### 3.1 Local alignment (Smith–Waterman) for near-repeated passages
- **Key citation.** Smith & Waterman, "Identification of common molecular subsequences", J. Mol. Biol. 147, 195–197 (1981) — UNVERIFIED.
- **Question answered.** Which pairs of lines (or windows) share a high-scoring local alignment allowing substitutions and gaps? Exact repeated phrases (already counted by the program) miss the near-repeats that both a copyist and a copy-and-modify generator produce.
- **Prior application.** Not applied to Voynich as alignment (search found nothing). Timm, "How the Voynich manuscript was created", arXiv:1407.6639 (2014) — VERIFIED (https://arxiv.org/abs/1407.6639) — argues from similar-word adjacency; Timm & Schinner, "A possible generating algorithm of the Voynich manuscript", Cryptologia 44(1), 1–19 (2020) — UNVERIFIED (only forum discussion surfaced) — formalises it. Alignment is the direct, quantitative test of their prediction.
- **Implementation.** Lines: on the order of 5,000. Affine-gap Smith–Waterman on U2 with a substitution matrix learned from the program's own edit-distance work (glyph pairs that frequently substitute get a smaller penalty; e.g. `o/a`, `r/s`, `ch/sh`, `in/iin`) and gap open/extend of −3/−1 in a +2 match scheme. All-pairs is 12.5M alignments of ~40×40 cells; in NumPy with a vectorised anti-diagonal kernel or numba this is ~1–2 h on 2 cores; alternatively prefilter with shared 4-mer counts (BLAST-style seeding) and align only the top 1% of pairs. Score significance with the Gumbel (Karlin–Altschul) fit to scores from N3-shuffled line pairs. Outputs: (i) distribution of best-hit scores per line vs controls; (ii) best-hit score as a function of manuscript distance between the two lines (in lines, pages, quires) — the Timm–Schinner model predicts strong locality (best match is on the same or adjacent page), a copied natural-language text does not; (iii) a near-duplicate graph (edges = significant alignments) with its connected components mapped onto folios. ~12 h including the substitution matrix and nulls.

### 3.2 Multiple sequence alignment of near-duplicate lines
- **Key citation.** Any progressive aligner (ClustalW/MAFFT-style) — UNVERIFIED.
- **Question answered.** For a cluster of near-duplicate lines, where are the conserved and variable columns, and are the variable columns the same glyph classes everywhere (which would argue for systematic variation, e.g. a cipher's nulls or homophones) or random?
- **Prior application.** Not applied.
- **Implementation.** Take the connected components from 3.1 with ≥ 3 members; build a progressive MSA with the same scoring; compute per-column conservation entropy; test whether variable positions are enriched for specific glyph classes against a null that permutes which columns vary. ~4 h on top of 3.1.

### 3.3 Median-joining networks of word variants
- **Key citation.** Bandelt, Forster, Röhl, "Median-joining networks for inferring intraspecific phylogenies", Mol. Biol. Evol. 16, 37–48 (1999) — UNVERIFIED.
- **Question answered.** For a family of word forms differing by single edits (e.g. `chedy / shedy / chey / chdy`), what is the network of minimal-edit paths, and is it star-like (one base form plus independent variants, as in a copyist's errors or a generator's modifications) or reticulate/chain-like (sequential derivation)?
- **Prior application.** Not applied to Voynich. The program's form–function tests already use edit distance vs context, which overlaps; the network *topology* and the mapping of variants to manuscript position do not.
- **Implementation.** Build the edit-distance-1 graph over U2 word types with frequency ≥ 3 (~2–3k nodes); extract families as connected components; compute per-family star-ness (fraction of nodes adjacent to the hub) and chain length; map each variant's first occurrence to folio order and test whether the hub form precedes variants (Spearman on first-occurrence order vs graph distance from hub). Controls: Latin spelling variation in a diplomatic transcription, and N6 output. PopART-style median-joining proper is unnecessary; the edit graph plus these statistics answers the question. ~8 h.

### 3.4 Textual stemmatics / phylogenetic methods
- **Key citations.** Barbrook, Howe, Blake, Robinson, "The phylogeny of The Canterbury Tales", Nature 394, 839 (1998), DOI 10.1038/29667 — VERIFIED (https://ideas.repec.org/a/nat/nature/v394y1998i6696d10.1038_29667.html). Howe et al., "Manuscript evolution", Trends in Genetics (2001) — VERIFIED title (https://www.academia.edu/1514562/Manuscript_evolution), details from memory. Painter & Bowern, "Examining the history of Voynich glyphs using phylogenetic methods", VOY2022 — VERIFIED listing (https://ceur-ws.org/Vol-3313/paper11.pdf).
- **Question answered.** Stemmatics needs multiple witnesses of one text; Voynich has one witness, so the classic application does not apply. Painter & Bowern used phylogenetic methods on glyph shapes vs comparison scripts, not on the text.
- **Prior application.** Applied (glyph level). Text-level stemmatics: not applicable without a second witness. The only within-manuscript analogue is 3.3.

---

## 4. Information-theoretic language detection

### 4.1 Zipf's law of abbreviation
- **Key citation.** Bentz & Ferrer-i-Cancho, "Zipf's law of abbreviation as a language universal", in *Proceedings of the Leiden Workshop on Capturing Phylogenetic Algorithms for Linguistics* (2016) — VERIFIED (https://publikationen.uni-tuebingen.de/xmlui/handle/10900/68639), venue from memory.
- **Question answered.** Is word length negatively correlated with word frequency? Found across ~1,000 languages by Bentz & Ferrer-i-Cancho (from memory); also found in some animal signal repertoires.
- **Prior application.** Not reported for Voynich in the collected material; Gaskell & Bowern's metric set includes word-length statistics but not this correlation explicitly (VERIFIED list of their metrics does not name it). **Not applied** as a formal test.
- **Implementation.** Kendall τ between log frequency and length over U3 types (and over U2-length, since EVA length inflates some glyphs), with a permutation p-value, for each section. Control values from Latin/Italian at equal token count; cipher controls matter here because vowel deletion and bigraph encoding both change length while preserving frequency rank. N6 output will also show the law if the generator copies frequent words preferentially, so this is a necessary-not-sufficient test. ~2 h.

### 4.2 Menzerath–Altmann law
- **Key citation.** Altmann, "Prolegomena to Menzerath's law", Glottometrika 2, 1–10 (1980) — UNVERIFIED. (The Torre et al. 2019 speech study I intended to cite did not surface; UNVERIFIED.)
- **Question answered.** Do longer constructs have shorter constituents (longer words → shorter "morphs"; longer lines → shorter words; longer paragraphs → shorter lines)? Holds in language at several levels and in some genomic and musical data.
- **Prior application.** Not applied to Voynich in the collected material.
- **Implementation.** Three levels: (word length in U2 units) vs (mean morph length from 2.5/2.6); (line length in words) vs (mean word length); (paragraph length in lines) vs (mean line length). Fit y = a·x^b·e^{−cx}; report b with bootstrap CI; compare to controls. The line level is confounded by physical line width (a long line is long because of the page), so the word→morph level is the only one that speaks to linguistic structure, and it depends on an induced segmentation — report it conditional on each of BPE, Morfessor and adaptor-grammar morphs. ~2 h once 2.5/2.6 exist.

### 4.3 Entropy rate vs vocabulary growth (Heaps' law)
- **Key citation.** Heaps, *Information Retrieval: Computational and Theoretical Aspects* (1978) — UNVERIFIED.
- **Question answered.** How fast does the type count grow with tokens (V ∝ N^β), and is that growth consistent with the entropy rate from 1.6? Natural languages pair β ≈ 0.5–0.7 with h ≈ 1–1.5 bits/char; a copy-and-modify generator can produce almost any β by tuning its modification rate.
- **Prior application.** Heaps behaviour is implicitly handled by Caruana et al. 2022 (random 10k-word segments to control for it; VERIFIED) and by Gaskell & Bowern's unique-word metrics. A joint (β, h) comparison: not applied.
- **Implementation.** Type–token curve by manuscript order and by 100 random orders; fit β on 1k–37k; plot (β, h) for Voynich A, B, controls, cipher controls, N6. ~2 h.

### 4.4 Burstiness and recurrence-interval statistics
- **Key citation.** Altmann, Pierrehumbert, Motter, "Beyond word frequency: bursts, lulls, and scaling in the temporal distributions of words", PLoS ONE 4(11): e7678 (2009) — VERIFIED (https://pmc.ncbi.nlm.nih.gov/articles/PMC2770836), article number from memory.
- **Question answered.** For each word type, is the distribution of gaps between successive occurrences Poisson (memoryless) or bursty (stretched exponential / Weibull with shape β < 1)? In language, content words are bursty and function words near-Poisson; burstiness tracks semantic class, not frequency.
- **Prior application.** Montemurro & Zanette 2013 extracted "keywords" via an entropy-of-distribution-over-sections measure (VERIFIED title; method from memory), which is a coarse-grained relative. Word-level inter-occurrence distributions: **not applied**.
- **Implementation.** For every U3 type with ≥ 20 tokens (a few hundred types), fit the Weibull shape β to its inter-occurrence intervals (in tokens) by MLE; also compute the Fano factor of counts in 500-token windows. Plot β against frequency and against the word's structural class (prefix `qo-`, suffix `-aiin`, `-edy`, gallows-initial). Null: N2 (global word shuffle) gives β = 1 for all; N7 preserves Zipf and long memory and shows what burstiness "memory alone" produces. Hypothesis: if Voynich words carry topic, β should spread well below 1 for a subset of mid-frequency words, independent of frequency, as in language; if words are generated by local copying, burstiness will be strong but *monotone in frequency* and concentrated at short lags (within a page), which the lag-resolved version of the test will show. ~6 h.

### 4.5 Word-length distributions and language-family comparison
- **Key citations.** Bowern & Lindemann, "The linguistics of the Voynich manuscript", Annual Review of Linguistics 7, 285–308 (2021) — VERIFIED PDF (https://alumniacademy.yale.edu/sites/default/files/2021-07/The%20Linguistics%20of%20the%20Voynich%20Manuscript.pdf), volume/pages from memory. Lindemann & Bowern 2020 (character entropy) — VERIFIED. Gaskell & Bowern 2022 — VERIFIED.
- **Prior application.** Applied: word-length distributions, their autocorrelation along the text, and character entropy, all compared against many languages and against human gibberish. Known results from the fetched material: Voynich word-length autocorrelation and Zipf fit are within the range of human gibberish; most Voynich transcriptions were classified as gibberish-like by their classifier.
- **Implementation.** Nothing new to add beyond re-running their 42-metric panel under the N7 null and the semi-Markov HMM states from 2.2.

### 4.6 Animal-communication tests
- **Key citations.** McCowan, Hanser, Doyle, "Quantitative tools for comparing animal communication systems: information theory applied to bottlenose dolphin whistle repertoires", Animal Behaviour 57, 409–419 (1999) — UNVERIFIED (search returned only secondary coverage). Ferrer-i-Cancho & McCowan, "The span of correlations in dolphin whistle sequences", J. Stat. Mech. P06002 (2012), DOI 10.1088/1742-5468/2012/06/P06002 — VERIFIED (fetched, https://arxiv.org/abs/1205.0321).
- **Question answered.** McCowan et al.: Zipf slope and higher-order Shannon entropies as "repertoire complexity" measures. Ferrer-i-Cancho & McCowan: how many previous symbols a symbol depends on, using global and local randomisation tests (they found dependence back to the 2nd previous whistle globally, 4th locally).
- **Prior application.** Not applied to Voynich as such; the "span of correlations" test is a cleaner version of what the program's MI decay already measures, with the distinction that the *local* randomisation (shuffling within a window) controls for non-stationarity.
- **Implementation.** On U3: for lag d = 1 … 10, compare I(X_t; X_{t−d}) against (a) global word shuffle and (b) shuffle within sliding windows of 50 words. The largest d at which Voynich beats the local null is its "dependency span"; compare with controls (language typically ≥ 5–8 words at 37k tokens; measure). ~2 h, mostly reuse.

### 4.7 SETI "detecting language" criteria
- **Key citation.** Doyle, McCowan, Johnston, Hanser, "Information theory, animal communication, and the search for extraterrestrial intelligence", Acta Astronautica 68, 406–417 (2011) — UNVERIFIED (search returned only coverage: https://archive.seti.org/..., https://www.cambridge.org/core/journals/symposium-international-astronomical-union/article/...).
- **Question answered.** Their criteria are the Zipf slope ≈ −1 and the decay of conditional entropies with order (the "entropic slope"). Both are already in hand for Voynich, and both are known to be satisfied by non-linguistic sources, so there is no new test here; the material matters for calibrating claims, not for new work.

### 4.8 Universal entropy of word ordering
- **Key citation.** Montemurro & Zanette, "Universal entropy of word ordering across linguistic families", PLoS ONE 6(5): e19875 (2011) — VERIFIED (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3094390/), article number from memory.
- **Question answered.** The difference between the entropy of a text and that of its word-shuffled version (the part of the entropy due to word order) is roughly constant (~3.5 bits/word, from memory) across language families.
- **Prior application.** Applied to Voynich by the same authors in 2013 (VERIFIED title). Re-running it against N7 rather than a plain shuffle would separate "order information" from "long-memory information"; ~2 h.

### 4.9 Uniqueness point / segment-information tests
- **Key citation.** Layfield, van der Plas, Rosner, Abela, "Word probability findings in the Voynich manuscript", LT4HALA 2020, pp. 74–78 — VERIFIED (fetched). Applied: less probable words have earlier uniqueness points, as in five natural languages. Nothing to add except running the same regression on N6 and cipher controls, which they did not do (~2 h).

---

## 5. Epigraphic and decipherment computing

### 5.1 Conditional entropy as a linguistic signature, and its critique
- **Key citations.** Rao, Yadav, Vahia, Joglekar, Adhikari, Mahadevan, "Entropic evidence for linguistic structure in the Indus script", Science 324, 1165 (2009) — VERIFIED as existing via Language Log discussion (https://languagelog.ldc.upenn.edu/nll/?p=1374), bibliographic details from memory. Sproat, "A statistical comparison of written language and nonlinguistic symbol systems", Language 90(2), 457–481 (2014) — VERIFIED (https://muse.jhu.edu/article/547119; https://research.google/pubs/a-statistical-comparison-of-written-language-and-nonlinguistic-symbol-systems/), volume/pages from memory. Rao et al. reply, "On statistical measures and ancient writing systems", Language (2015) — VERIFIED (https://resolve-he.cambridge.org/core/journals/language/article/on-statistical-measures-and-ancient-writing-systems/D75CB72F35D752E4848CA69AE8E32888).
- **Question answered.** Rao et al. argued that the bigram conditional entropy of the Indus signs falls in the band occupied by natural languages, between rigid (DNA-like, low) and random (high) systems. Sproat showed, with real corpora of nonlinguistic symbol systems (heraldry, Pictish symbols, Mesopotamian deity symbols, Vinča etc., from memory) that the conditional-entropy band does not discriminate, and that Rao's low/high extremes were artificial.
- **Prior application.** Character conditional entropy for Voynich is well documented (Lindemann & Bowern 2020, VERIFIED; it is anomalously low). The methodological point transfers directly: a single entropy statistic cannot establish linguisticity; what moved opinion in the Indus debate was the *comparison set* (real nonlinguistic corpora), and no Voynich study has used nonlinguistic symbol corpora as controls. Gaskell & Bowern's human gibberish is the nearest thing.
- **Implementation.** Assemble two or three nonlinguistic sequential-symbol corpora in the public domain (e.g. heraldic blazon tokens, a tokenised musical score corpus, a chess-move corpus) and add them as a control family to the standard panel. This is a data-assembly task (~6 h) rather than an algorithm, but it is the single most-cited gap in "is it language" arguments.

### 5.2 n-gram Markov modelling of an undeciphered script (Indus)
- **Key citation.** Yadav, Joglekar, Rao, Vahia, Adhikari, Mahadevan, "Statistical analysis of the Indus script using n-grams", PLoS ONE 5(3): e9506 (2010) — VERIFIED (https://pmc.ncbi.nlm.nih.gov/articles/PMC2841631; arXiv:0901.3017).
- **Question answered.** Which signs are text-initial/terminal, which bigrams are significant, and can a smoothed n-gram model fill in damaged positions? Essentially a positional-grammar study.
- **Prior application.** The Voynich equivalents exist (Feaster 2022; Reddy & Knight 2011; quimqu's n-gram perplexity; Parisel 2025 directionality via n-gram perplexity asymmetry, VERIFIED https://arxiv.org/abs/2509.10573). The Indus "fill in the blank" use is the under-tried piece: predicting masked tokens as a way of scoring how deterministic the text is. See 6.1.

### 5.3 Linear A
- **Key citation.** Salgarella & Castellan, "SigLA: The signs of Linear A. A palaeographical database" — VERIFIED (https://sigla.phis.me/paper.html; https://www.repository.cam.ac.uk/items/25c5d712-6e4c-4435-86a5-30f769ea4072), authorship from memory.
- **Relevance.** A sign-level palaeographic database with per-attestation shape variants, which is the infrastructure that lets sign-frequency and allograph statistics be computed reproducibly. Voynich has no equivalent public glyph-instance database (Davis's work used Archetype, 7.1); building one is a prerequisite for several image-side methods and is out of scope for a text-sequence pipeline.

### 5.4 Rongorongo
- **Key citations.** Horley, "Allographic variations and statistical analysis of the rongorongo script", Rapa Nui Journal — VERIFIED title (https://evols.library.manoa.hawaii.edu/items/af73d1dc-60df-4143-ab68-770a56d81d82/full), year from memory (2005). Horley, "Structural analysis of rongorongo inscriptions", Rapa Nui Journal — VERIFIED title (https://evols.library.manoa.hawaii.edu/items/70111ccf-3ead-459a-854f-b7347d7ac6d3/full), year from memory (2007). Pozdniakov & Pozdniakov, "Rapanui writing and the Rapanui language: preliminary results of a statistical analysis" — VERIFIED as likely https://hal.archives-ouvertes.fr/hal-01389639 (title not shown in result). Lastilla, Ravanelli, Valério, Ferrara, "Modelling the rongorongo tablets", Digital Scholarship in the Humanities 37(2), 497–516 (2022), DOI 10.1093/llc/fqab045 — VERIFIED (fetched).
- **Methods transferable.** Pozdniakov's main move (from memory) was to reduce the sign inventory by merging allographs until sign-frequency and sign-sequence statistics matched Rapanui syllable statistics; Horley's was structural analysis of repeated sequences and parallel passages across tablets (an alignment problem, cf. 3.1). The rongorongo "parallel passages" work is the closest epigraphic precedent for applying local alignment to Voynich.

### 5.5 Proto-Elamite
- **Key citation.** Born, Kelley, Kambhatla, Chen, Sarkar, "Sign clustering and topic extraction in Proto-Elamite", ACL LaTeCH-CLfL 2019, W19-2516 — VERIFIED (https://aclanthology.org/W19-2516/); author list partly from memory (Kelley and Chen seen). Also an SFU thesis, "Applications of natural language processing to archaeological decipherment: a survey of Proto-Elamite" — VERIFIED (https://summit.sfu.ca/_flysystem/fedora/2023-12/etd22782.pdf).
- **Methods transferable.** Sign embeddings from co-occurrence, clustering to propose allograph merges, and topic models over tablets. The Voynich analogue (topic models) exists: Sterneck, Polish, Bowern, "Topic modeling in the Voynich manuscript", arXiv:2107.02858 (2021) — VERIFIED (https://arxiv.org/abs/2107.02858v1). Embedding-based allograph merging on EVA (asking which EVA characters behave as contextual variants of one another) is under-tried; Zattera's VOY2022 transliteration-alphabet paper approaches it manually (VERIFIED listing, https://ceur-ws.org/Vol-3313/paper10.pdf).

### 5.6 Etruscan
- Not covered by the collected material; UNVERIFIED and omitted rather than invented. (Etruscan is readable, not undeciphered, so its computing literature is mostly lexicography.)

### 5.7 What statistics convinced people that a sequence is linguistic
- From the collected material, the honest answer is: none, on their own. Rao's entropy argument was neutralised by Sproat's control corpora; Zipf fits were neutralised in the DNA debate; Gaskell & Bowern 2022 showed human gibberish matches Voynich on most low-level metrics; Bowern & Gaskell 2022 showed several ciphers match it too. Decipherments that did convince (Linear B, Maya — UNVERIFIED general background) rested on structural paradigms (Kober's inflectional triplets) and external anchors, not on entropy. The methodological implication for ranking: prefer tests whose outcome distinguishes *generating mechanisms* (copy-and-modify vs cipher vs language) rather than tests of "language-likeness".

---

## 6. Deep learning and representation learning on unlabeled sequences

### 6.1 Character-level language-model perplexity as a scoring device
- **Prior application.** quimqu (Voynich Ninja, June 2025) trained 1- to 9-gram character models on EVA and comparison corpora and found Voynich perplexity unusually low at high orders, closer to medieval religious/philosophical texts than to literary ones; commenters noted the EVA-encoding dependence (VERIFIED, https://www.voynich.ninja/thread-4763.html). Parisel 2025 used n-gram perplexity asymmetry to test reading direction (VERIFIED). Turenne, arXiv:2609.20835 (2026), uses position-dependent probabilistic grammars and LLM-based image–text alignment; the fetched abstract does not report perplexity (VERIFIED).
- **Under-tried.** A cross-section perplexity matrix with a *neural* character model and the glyph-merged alphabet, with proper held-out evaluation: train on each of {Herbal-A, Herbal-B, Bio, Stars, Pharma, Astro, each scribe} and test on each of the others. Natural-language texts in one language give a near-flat matrix (small train/test mismatch); a text in two "languages" or two cipher keys gives a block structure; copy-and-modify text gives a matrix that depends on manuscript distance.
- **Implementation.** A 2-layer LSTM or a 4-layer, 128-dim transformer with context 256 on U2 trains in well under an hour per section on 2 CPU cores with PyTorch (the corpus is ~190k symbols); use 5-fold line-level cross-validation inside each section to get the diagonal; report bits-per-symbol with bootstrap CIs. Add the Indus-style masked-token task: mask 10% of word tokens, predict with a word-level model trained on the rest, and report accuracy vs controls — a measure of how deterministic word choice is. ~8 h.

### 6.2 Word embeddings (word2vec / FastText) on Voynich
- **Prior application.** Lhotka, Salvi, Lazauskas (EPFL, supervised by van der Plas), "Disambiguating Voynich manuscript transliterations with word embeddings" — VERIFIED (fetched, https://www.epfl.ch/labs/mlo/wp-content/uploads/2022/10/crpmlcourse-paper1185.pdf): FastText with softmax ranking resolved artificial Voynich-style ambiguities in Italian/Latin at ~86% vs ~50% baseline; word2vec performed poorly at this corpus size. Perone's blog post on word vectors and t-SNE of Voynich — VERIFIED (https://blog.christianperone.com/?p=2974).
- **Assessment.** At 37k tokens, word embeddings are noisy; FastText's subword sharing is the only reason it works, and that makes its "semantics" largely orthographic. Useful as a transliteration-cleaning tool (their application), not as a semantic probe. ~4 h to reproduce; low priority.

### 6.3 Unsupervised cross-lingual embedding alignment (MUSE)
- **Key citations.** Conneau, Lample, Ranzato, Denoyer, Jégou, "Word translation without parallel data", ICLR 2018, arXiv:1710.04087 — VERIFIED (https://arxiv.org/abs/1710.04087v1). Lample et al., "Unsupervised machine translation using monolingual corpora only", ICLR 2018, arXiv:1711.00043 — VERIFIED (https://ar5iv.arxiv.org/html/1711.00043).
- **Question answered.** Can a linear map align two monolingual embedding spaces with no dictionary (adversarial + Procrustes refinement)? Works for large, related, same-domain corpora; degrades sharply for small or distant corpora (from memory of follow-up literature).
- **Prior application.** Not applied to Voynich in the collected material. Given 6.2's finding that even monolingual embeddings are weak at this size, MUSE alignment to Latin/Hebrew/etc. is expected to fail for reasons unrelated to the manuscript, and a failure would be uninterpretable. **Not recommended** without a positive control showing MUSE aligns a 37k-token Latin corpus to a 37k-token Italian corpus at all (that control alone is ~4 h and is where to start if anyone insists).

### 6.4 Neural decipherment with cognate/phonetic priors
- **Key citations.** Luo, Cao, Barzilay, "Neural decipherment via minimum-cost flow: from Ugaritic to Linear B", ACL 2019, arXiv:1906.06718 — VERIFIED (https://arxiv.org/abs/1906.06718). Luo, Hartmann, Santus, Barzilay, Cao, "Deciphering undersegmented ancient scripts using phonetic prior", TACL 9 (2021) — VERIFIED (https://transacl.org/index.php/tacl/article/view/2449), authorship partly from memory.
- **Question answered.** Given an unknown script and a *known related language*, learn a character mapping by minimum-cost-flow matching of embeddings under a monotone-alignment constraint. The 2021 paper handles unsegmented text and ranks candidate known languages by decipherment loss (they applied it to Iberian, from memory).
- **Prior application.** Not applied to Voynich in the collected material. Hauer & Kondrak, "Decoding anagrammed texts written in an unknown language and script", TACL 4 (2016) — VERIFIED (https://aclanthology.org/Q16-1006/) — is the Voynich precedent: a language-identification plus anagram-decipherment pipeline that ranked Hebrew highest (from memory).
- **Implementation.** The Luo 2021 "language ranking by decipherment loss" is the usable piece: run the released code against a panel of 20 candidate languages with the Voynich U2 stream, but *first* run the same panel on the cipher controls (vowel-deleted Latin, bigraph-encoded Latin) to see whether the method can recover the true source language under those manipulations. If it cannot, Voynich results are uninterpretable. ~16 h; ranked low because of this dependency.

### 6.5 LLM-based approaches
- Turenne 2026 (VERIFIED) is the only such paper in the material; its text-side claims rest on probabilistic grammars and Zipf fits, and the LLM component is image–herbal alignment. Nothing to rank.

---

## 7. Image and diagram analysis in manuscript studies

### 7.1 Computational palaeography / scribal hands
- **Key citation.** Davis, "How many glyphs and how many scribes? Digital paleography and the Voynich manuscript", Manuscript Studies 5(1) (2020) — VERIFIED (https://muse.jhu.edu/article/754633; https://repository.upenn.edu/mss_sims/vol5/iss1/6). Davis used the Archetype platform to annotate glyph instances and identified five scribal hands (from memory).
- **Relevance to the text pipeline.** The scribe assignment is metadata every test above should condition on; several of the ranked items (1.7, 2.3, 6.1) test whether statistical structure follows scribe rather than section. Farrugia, Layfield, van der Plas, "Demystifying the scribes behind the Voynich manuscript using computational linguistic techniques", VOY2022 — VERIFIED listing (https://ceur-ws.org/Vol-3313/) — is the text-side counterpart.

### 7.2 Glyph clustering / OCR
- No "VoynichOCR" paper surfaced in the searches; UNVERIFIED and treated as not existing in print. The Linear A SigLA database (5.3) shows what the output should look like. Not ranked (image work, outside a 37k-token sequence pipeline).

### 7.3 Multispectral imaging
- Reported September 2024: multispectral scans of f1r revealed an erased earlier decoding attempt/ownership marks (coverage VERIFIED: https://www.medievalists.net/2024/09/hidden-secrets-of-the-voynich-manuscript-revealed-through-cutting-edge-scans/; https://theartnewspaper.com/2024/09/25/voynich-manuscript-scans-reveal-early-decoding-attempt). Pigment analysis from photographs: not in the material; UNVERIFIED and omitted.

### 7.4 Layout analysis
- Feaster 2022 (rightward/downward grapheme distributions) and Stafford 2022 ("Seven habits of highly eccentric paragraphs") — VERIFIED listings — are layout-statistics papers on the transliteration. Line-position effects are strong enough that every sequence test must either condition on position or include N3/N4 nulls; this is already built into the conventions in section 0.

### 7.5 Graph representations of diagrams
- Not in the collected material; no Voynich application found. UNVERIFIED and not ranked. (Label text on diagrams vs paragraph text is a *text* question that can be run with the existing pipeline: compare label-token statistics — length, prefix distribution, burstiness — against paragraph tokens; ~2 h, included as a sub-step of item 9 below.)

---

## 8. Ranked top-10 untried or under-tried techniques

Excluded from ranking because already run by the program: MI decay, gzip compressibility, BPE merge curves, repeated-phrase counts, 2/3-state HMM, form–function edit-distance-vs-context tests. Ranking criterion: probability that the result discriminates between the three live generating hypotheses (natural language possibly in an unusual orthography; cipher of a natural language; copy-and-modify or other meaningless generation) given a 37k-token corpus, weighted by how cheaply the result can be made defensible with the nulls and controls in section 0. Hours are for Python on 2 cores, including nulls, controls and a write-up-ready figure set.

**1. Local alignment sweep for near-duplicate passages (3.1, with 3.2).** Hypothesis: the Timm–Schinner copy-and-modify process predicts that each line's best local-alignment partner lies within a few lines or on the same page, with scores far above the Gumbel null; a copied or composed natural-language text (and its ciphers) predicts weak locality and near-repeats only for formulaic phrases. This is the first test that directly operationalises the self-citation hypothesis at the passage level rather than the word level, it reuses the program's edit-distance work for the substitution matrix, and its output (a near-duplicate graph over folios) is a reusable asset for items 4 and 5. The main risk is that EVA transcription errors inflate near-repeats, which the control (a diplomatic transcription of a Latin herbal with its own scribal variation) addresses. **12 h.**

**2. Bayesian word segmentation of the de-spaced text (2.4).** Hypothesis: if scribal spaces mark lexical units, an HDP-bigram learner recovers them at the same F1 it achieves on de-spaced Latin/Italian; if Voynich "words" are composites of smaller recurring units (cipher groups, slot fillers), the learner systematically splits them at the same internal seams (`qo|`, `|aiin`, `|edy`), and if words are produced by copy-and-modify, the learner recovers spaces *more* easily than in real language because word edges are highly predictable. Three distinct outcomes map onto the three hypotheses, which is why this ranks high. Off-the-shelf samplers exist; most of the effort is controls and the uncertain-space handling. **10 h.**

**3. Zipf-preserving long-range surrogate as the null for the existing word-level battery (1.9, with DFA 1.3).** Hypothesis: several of the program's existing positive findings (MI decay, repeated-phrase excess, burstiness if added) are reproduced by a surrogate that preserves only word frequencies and the DFA exponent, in which case they say nothing beyond "Zipf plus memory"; whatever survives is structure that neither plain shuffles nor long memory explains. This upgrades the defensibility of everything already done at modest cost, and it brings DFA to Voynich in its modern form with proper confidence intervals, replacing the 2001 Hurst/spectral result. **10 h** (6 h surrogate and re-runs, 4 h DFA with crossover analysis).

**4. Sequitur / MDL grammar statistics (2.8).** Hypothesis: natural language yields a deep, globally reused rule hierarchy whose rule-use spans are spread over the whole manuscript; copy-and-modify yields shallow rules with short spans concentrated within pages; ciphers that preserve word identity look like language, while vowel-deleted or bigraph-encoded ciphers change rule depth in measurable ways that the controls calibrate. The "rule span" statistic is cheap and, to my knowledge from the collected material, has never been computed for Voynich. Runs on both characters and word IDs with the same code. **8 h.**

**5. Line-level multivariate change-point detection aligned with scribes and quires (2.3).** Hypothesis: the statistical breaks in the text (prefix/suffix shares, gallows rate, A/B score, word length) coincide with Davis's scribe changes and quire boundaries more than with illustration-section boundaries; the published folio-level, two-feature Bayesian analysis found breaks around f33/f38 but did not test this alignment. A positive result would tie text statistics to production rather than content, which bears on all three hypotheses; a negative result (breaks follow content sections) supports the text having topic structure. PELT in `ruptures` is fast; the effort is in building the feature matrix and the alignment scoring. **8 h.**

**6. Semi-Markov / many-state HMM with word-level emissions and duration (2.2b).** Hypothesis: the text contains runs of words with characteristic "type" and characteristic run length (the gene-finder exon/intron analogue); under language, state durations are broad and states correspond to syntactic/topical mixtures; under copy-and-modify, states collapse to "recently copied vs fresh" with durations tied to the modification schedule, and under a positional cipher, states align with line position. Model selection over K = 2…12 on the character level is the cheap first half and extends the program's 2/3-state runs; the semi-Markov word model is the new part. **10 h.**

**7. Cross-section neural character-LM perplexity matrix with the glyph-merged alphabet, plus masked-token prediction (6.1).** Hypothesis: perplexity off the diagonal of a section×section (and scribe×scribe) matrix reveals whether A and B are two systems (block structure) or one system with drift (distance-dependent), and whether the anomalously low character entropy is an EVA artefact (it should shrink substantially under U2 if so, as forum critics argued). Masked-word prediction accuracy, borrowed from the Indus n-gram work, gives a single number for word-choice determinism that can be compared across controls. Small models train in minutes on CPU; the cost is the cross-validation grid. **8 h.**

**8. Unsupervised morphology induction with MDL comparison (2.5 + 2.6, feeding 4.2).** Hypothesis: Morfessor and adaptor-grammar cost curves and morph inventories for Voynich fall outside the envelope of Latin/Italian/German controls in a specific direction (very few, highly reusable "morphs" with near-positional distribution), and the three segmenters (BPE already run, Morfessor, adaptor grammar) agree on seams far more than they do on real language, which would indicate a slot grammar rather than concatenative morphology. Also delivers the morph level required for a Menzerath–Altmann test. Mostly off-the-shelf tools. **8 h.**

**9. Burstiness and recurrence-interval statistics per word type, including label vs paragraph text (4.4, with 7.5's text sub-step).** Hypothesis: in language, Weibull shape β of inter-occurrence intervals spreads well below 1 for a subset of mid-frequency words independently of frequency (topic words); under copy-and-modify, burstiness is strong but monotone in frequency and concentrated at within-page lags; under cipher, the language pattern survives. The lag-resolved version is what makes this discriminating, and the N7 surrogate from item 3 is the right null. Adding the label-token vs paragraph-token comparison costs little and tests whether labels behave like a different register (names) as they would in a herbal. **6 h.**

**10. Zipf's law of abbreviation and Menzerath–Altmann across levels (4.1, 4.2).** Hypothesis: Voynich shows the frequency–length anticorrelation at the strength seen in languages and in vowel-deleted ciphers, and a Menzerath relation between word length and induced-morph length; copy-and-modify output shows the abbreviation law only if the generator copies frequent words preferentially, which can be checked on N6 directly. Cheap, well-defined, with a large cross-linguistic reference literature; ranked last only because both laws are known to be weakly discriminating on their own, so their value is as part of the panel rather than as a standalone finding. **4 h** (2 h each, given items 3 and 8).

Not ranked, with reasons: NCD clustering (1.7) and zipping classification (1.8) are quick (3–4 h) but mainly reproduce what the perplexity matrix in item 7 does better; MEME/Gibbs motif finding (2.1) is worth 6 h once items 1 and 5 have shown where positional structure lives; median-joining variant networks (3.3) overlap substantially with the form–function tests already run; MUSE alignment (6.3) and neural decipherment (6.4) require a positive control at 37k tokens that the collected evidence (EPFL embeddings study) suggests will fail; nonlinguistic control corpora (5.1) are a data-assembly task that should be scheduled alongside whichever ranked item is run first, since they strengthen all of them.

---

## Citation index with verification status

VERIFIED (URL/DOI seen): Montemurro & Zanette 2013 (PMC3689824); Amancio et al. 2013 (PMC3699599, arXiv:1303.0347 — applied complex-network and intermittency statistics to Voynich, summary from memory); Bowern & Lindemann 2021 (Yale PDF); Gaskell & Bowern 2022 (CEUR paper4); Bowern & Gaskell 2022 (paper6); Caruana, Layfield, Abela 2022 (paper8); Layfield et al. 2020 (LT4HALA); VOY2022 volume listing (CEUR Vol-3313) incl. Lindemann, Zattera, Painter & Bowern, Feaster, Stafford, Farrugia et al., Boxer; Hauer & Kondrak 2016 (Q16-1006); Reddy & Knight 2011 (W11-1511); Lindemann & Bowern 2020 (arXiv:2010.14697); Sterneck, Polish, Bowern 2021 (arXiv:2107.02858); Davis 2020 (MUSE 754633); Timm 2014 (arXiv:1407.6639); Arutyunov et al. 2016 (arXiv:1611.09122); Parisel 2025 (arXiv:2509.10573); Turenne 2026 (arXiv:2609.20835); hepialidae 2020 change-point post; quimqu 2025 n-gram thread; Lhotka, Salvi, Lazauskas EPFL embeddings paper; Perone blog; Montemurro & Degli Esposti 2026 (arXiv:2603.02213); Montemurro & Zanette 2011 (PMC3094390); Altmann, Pierrehumbert, Motter 2009 (PMC2770836); Ferrer-i-Cancho & McCowan 2012 (arXiv:1205.0321, DOI 10.1088/1742-5468/2012/06/P06002); Takahira, Tanaka-Ishii, Dębowski 2016 (MDPI Entropy 18/10/364; W16-4124); Cilibrasi & Vitányi 2005 (CWI PDF, arXiv:cs/0312044); Benedetto, Caglioti, Loreto 2002 (arXiv:cond-mat/0108530) and Goodman's comment (cond-mat/0202383) and reply (cond-mat/0203275); Bonhoeffer et al. comment (adap-org/9503002); Niyogi & Berwick note (cmp-lg/9503012); Peng et al. 1992 and 1994 (titles via cris.biu.ac.il); Goldwater, Griffiths, Johnson 2009 (Princeton PDF); Nevill-Manning & Witten 1997 (arXiv:cs/9709102, JAIR); Barbrook et al. 1998 (Nature, DOI 10.1038/29667 via ideas.repec); Howe et al. 2001 (title); Luo, Cao, Barzilay 2019 (arXiv:1906.06718); Luo et al. 2021 (TACL 2449); Yadav et al. 2010 (PMC2841631, arXiv:0901.3017); Sproat 2014 (MUSE 547119); Rao et al. 2015 reply (Cambridge Language); Rao et al. 2009 (existence via Language Log p=1374); Horley rongorongo papers (eVols titles); Lastilla et al. 2022 (DOI 10.1093/llc/fqab045); Born/Kelley et al. 2019 Proto-Elamite (W19-2516); SigLA (sigla.phis.me); Eskander et al. 2016 (C16-1086); Creutz & Lagus 2007 (ufal PDF); Bentz & Ferrer-i-Cancho 2016 (Tübingen repository); Conneau et al. 2018 (arXiv:1710.04087); Lample et al. 2018 (arXiv:1711.00043); 2024 multispectral coverage (medievalists.net, The Art Newspaper); Gaskell's GitHub README.

UNVERIFIED (from memory; no direct hit in the collected material or the six extra searches): Landini 2001 Cryptologia; Timm & Schinner 2020 Cryptologia; Schinner 2007 Cryptologia (only 2007 news coverage seen: https://boingboing.net/2007/04/17/analysis-suggests-th.html); Mantegna et al. 1994 PRL (only the comments seen); Li 1990; Li & Kaneko 1992; Lempel & Ziv 1976; Willems, Shtarkov, Tjalkens 1995; Smith & Waterman 1981; Bandelt, Forster, Röhl 1999; Bailey & Elkan 1994; Lawrence et al. 1993; Killick, Fearnhead, Eckley 2012; Johnson, Griffiths, Goldwater 2007; Virpioja et al. 2013; Altmann 1980 (Menzerath); Heaps 1978; Burge & Karlin 1997; McCowan, Hanser, Doyle 1999; Doyle et al. 2011; Pozdniakov & Pozdniakov 2007 (HAL record seen without title); Stolfi's crust–mantle–core web pages; any "VoynichOCR"; Kober/Ventris background. Volume, page and DOI details attached to VERIFIED entries but marked "from memory" in the text should be checked before publication.