Research

EVA (European Voynich Alphabet)

Created by Rene Zandbergen and Gabriel Landini (1998). Defines ~25 basic characters using ASCII equivalents. 'Analytical' approach that breaks visible strokes into components (e.g., 'ch' = c + h). Most widely used system for computational analysis.

Gallows Characters

Four tall, ornate characters (EVA: t, k, p, f) that extend above the line. Appear mainly at word beginnings and line beginnings. 'Rare' gallows (cth, ckh, cph, cfh) combine bench + gallows. Their distribution suggests they may be numerals, abbreviation markers, or paragraph markers.

The 'Slot' Model (Stolfi)

Voynich words follow a rigid crust-mantle-core structure with specific character subsets at each layer. This grammar covers 96.5% of all running text tokens. Word length distribution is unusually binomial (peaked, not the long tail seen in natural languages). Position 1: q, s, d (or empty). Position 2: o (or empty). Position 3: l, r (or empty). Core: e, ch, sh, etc. Final: y, m, n, g. Unlike any known natural language or cipher.

Currier A vs B

Two statistically distinct 'languages' first identified by Prescott Currier (1976). Language A: more 'o' characters, found in herbal-A sections. Language B: more 'a' characters, found in herbal-B and biological sections. Could represent two dialects, scribes, or encoding schemes.

Entropy Anomalies

Character-level entropy: ~4.5 bits (similar to English). Word-level entropy: ~10 bits/word (lower than most European languages at ~11-12). Second-order entropy (character pairs) is unusually low, suggesting high predictability. This is the 'too regular' problem — the text is more predictable than natural languages.

Zipf's Law Compliance

Word frequencies follow Zipf's law (frequency inversely proportional to rank) very closely, matching natural language behavior. This was initially cited as evidence of real language, but it's now known that some generated texts can also exhibit Zipf-like distributions.

Hapax Legomena

Words appearing only once make up 14-20% of the vocabulary (depending on transcription). This is lower than typical natural languages (~40-60%) but higher than random text. The rate varies significantly based on which transliteration alphabet is used — a key insight from Zandbergen (2022).

Line as Functional Unit

First words on each line are ~1 character longer than average. Certain words have strong line-position affinity — one character appears line-finally in 85% of its occurrences. No known natural language shows this property. Lines may function as encoding units rather than arbitrary text wrapping.

Labels vs Paragraph Text

Labels (text near illustrations) have 12.4% 'abnormal' words vs only 3.7% in running text. Labels show very few repetitions and many unique forms. Interestingly, many label words also appear in running text — but not near their labeled illustration, suggesting labels aren't simple captions.

Extreme Positional Constraints

Characters show extreme positional preferences: 'q' appears only word-initially, 'm' only word-finally, 'y' only at word start or end. Bigram contact rules are far tighter than any natural alphabet — certain character pairs never occur despite both being common individually.

No Corrections

The manuscript shows virtually no corrections, cross-outs, or scribal errors. This is unusual for a 15th-century manuscript and has been cited both as evidence of careful copying (real text) and as evidence of meaningless generation (hoax). Genuine manuscripts of this era typically show corrections.

The 'Double Unknown' Problem

The Voynich presents a unique challenge: both the cipher method AND the underlying language are unknown. In every historical decipherment success (hieroglyphs, Linear B, Enigma), at least one was known. This 'double unknown' means the search space is combinatorially vast.

Identified 166 phytomorphs as New World species in their 2019 book 'Flora of the Voynich Codex' (Springer). Includes sunflower, chili pepper, armadillo. Won the American Botanical Council's James A. Duke Excellence in Botanical Literature Award (2020).

Award-winning botanical analysis but conflicts with radiocarbon dating (vellum 1404-1438, predating European knowledge of New World). Tucker argues text was written later on old vellum. Mainstream scholars remain skeptical of the New World hypothesis.

Identified several plants as common European species, suggesting Northern Italian origin. Compared illustrations to works by Leonardo da Vinci's contemporaries.

Some identifications are reasonable but many botanical illustrations appear to be composites or fantasy plants, making identification speculative.

Used botanical identifications as entry points for decipherment. Matched illustration of Centaurea (cornflower) to the label text, proposing phonetic values for several characters.

Method is sound in principle (similar to how Egyptian hieroglyphs were deciphered via bilingual texts), but sample size too small for statistical validation.

Diffusion model learned to transform ancient Oracle Bone characters into modern Chinese. Won ACL 2024 Best Paper. Follow-ups (DCSD-OBI, OracleFusion at ICCV 2025) improved accuracy by 11%+. Uses Chinese-CLIP for cross-modal consistency.

Relevance: Direct precedent: train a diffusion model on Voynich glyphs to learn structural decomposition and morph toward known alphabets.

Transformer trained on 78,000 ancient Greek inscriptions. 62% accuracy on text restoration alone; 72% when combined with historians. Geographic attribution at 71% accuracy. Dating within 30 years. Published in Nature.

Relevance: Shows self-supervised pre-training works on ancient corpora. Our 37,919 word tokens may be sufficient for character-level pattern learning.

CT scanning + AI read 2,000+ characters from carbonized scrolls where ink was invisible to the eye. $700K grand prize awarded February 2024. Discovered a previously unknown tract by Philodemus.

Relevance: Proof that AI can extract information from manuscripts that humans literally cannot see. Multispectral + AI pipeline could reveal hidden Voynich features.

Minimum-cost flow optimization deciphered Ugaritic (via Hebrew) and Linear B (via Greek). 2021 follow-up handled undersegmented scripts with phonetic priors. Published ACL 2019 and TACL 2021.

Relevance: Best algorithmic framework for cipher-breaking, but requires a known related language — the Voynich's fundamental blocker.

Combinatorial optimization tested on Linear A, Proto-Elamite, Indus Valley, Rongorongo. Allows null, one-to-many, and many-to-one character mappings. Published Frontiers in AI.

Relevance: Could be applied directly to Voynich under various language hypotheses. Handles verbose mappings naturally.

42 volunteers wrote meaningless text. ML classifier found Voynich transcriptions statistically resemble human-produced gibberish more than meaningful text. Code on GitHub (danielgaskell/voynich).

Relevance: Serious challenge: any decipherment theory must explain why the text passes this gibberish test. Counter: Montemurro's semantic clustering is harder to fake.

Treats Voynich as an encoded bitstream, testing decompression parameters using Shannon entropy as fitness. Proposes low redundancy and word structure are artifacts of compression (LZ77 + Huffman-like).

Relevance: Novel framework. If the text is compressed rather than encrypted, the decipherment problem changes fundamentally.

Open-source Python implementation (github.com/greshko/naibbe-cipher). Includes encryption, decryption, and Voynich-style text generator. Verbose homophonic substitution using dice + playing cards.

Relevance: Concrete, testable cipher model. We can generate synthetic Naibbe ciphertexts and compare statistics to the real manuscript systematically.

Multiple benchmark studies (CipherBench, CipherBank at ACL 2025) tested GPT-4, Claude, and Gemini on classical cipher-breaking. LLMs can break simple substitution ciphers (Caesar, Atbash) with ~90% accuracy but fail on polyalphabetic and homophonic ciphers. Performance degrades sharply with cipher complexity.

Relevance: Establishes that current LLMs cannot brute-force the Voynich directly. But LLMs excel at pattern description and hypothesis generation — use them as analytical partners, not as cipher-breakers.

Kambhatla et al. (EMNLP 2018) replaced n-gram language models with LSTMs in hill-climbing cipher attacks, achieving 2x improvement on homophonic substitution ciphers. Extended by beam search decipherment (ACL 2022). The same technique powered AZdecrypt's solution of the Zodiac Z340 cipher in 2020.

Relevance: This is the core technique for our Naibbe attack strategy. Neural LM scoring of candidate plaintexts is dramatically better than classical n-gram methods.

David Oranchak, Jarl Van Eycke, and Sam Blake used AZdecrypt to crack the 51-year-old Zodiac Z340 cipher. The tool uses hill climbing with n-gram scoring, testing 650,000+ candidate transpositions. FBI confirmed the solution in March 2021.

Relevance: Closest real-world precedent for a computational Voynich attack. Z340 was a homophonic substitution cipher with transposition — similar complexity class. The Voynich has 500x more ciphertext, giving more statistical leverage.

Ferrara (PNAS 2022) applied unsupervised neural embeddings to cluster signs in the undeciphered Cypro-Minoan script. Learned representations captured structural relationships without any labeled data. Identified potential sign variants and ligatures.

Relevance: Directly applicable to visual alphabet discovery. Train embeddings on our 17,672 glyph images to discover the 'true' character set from pixels alone, bypassing transcription assumptions.

Gutherz et al. (PNAS Nexus 2023) built an NMT system translating Akkadian cuneiform to English. Trained on ~10,000 parallel text pairs. Companion model 'cuneiformBase-400m' handles multiple ancient languages. Follow-ups (ICLR 2025) added iterative translation refinement.

Relevance: Shows NMT works on ancient languages with sufficient parallel data. The Voynich lacks parallel text — but if even partial decipherment produces fragments, NMT could bootstrap the rest.

GPT-4V and Claude Vision applied to medieval manuscript analysis — page layout detection, script identification, illustration classification. HTR-United and CATMuS Medieval provide training data (200+ manuscripts, 160K+ lines). Florence-2 and Grounding DINO enable zero-shot object detection in manuscript pages.

Relevance: We already used Claude Vision for glyph extraction (17,672 glyphs). Next: systematic page-level analysis combining text regions, illustration types, and marginalia to build a structural map of the entire manuscript.

The Joint Segmentation + Decryption Problem

Unlike standard homophonic ciphers (1 symbol = 1 letter), a verbose cipher requires solving TWO problems simultaneously: (1) how to segment Voynich "words" into letter-encoding units (unigrams vs bigrams), and (2) what each unit decrypts to. This makes it fundamentally harder than Z340 — but the Voynich has ~170,000 characters (vs Z340's 340), giving far more statistical leverage.