Unraveling the history, theories, and a data-driven attack on the world’s most mysterious codex
The Voynich Manuscript (Beinecke MS 408) is a ~240-page illustrated codex, carbon-dated to 1404–1438, written in an unknown alphabet. For over a century it has defeated the best codebreakers, including the founders of modern cryptanalysis.
The codex measures 22.5 × 16 cm and comprises roughly 240 vellum pages, with missing pages indicated by gaps in the 17th-century foliation. The text is written left-to-right in brown iron-gall ink with washes of green, blue, red, and yellow.
Physical evidence confirms the manuscript is a 15th-century artifact, ruling out a modern hoax.
The illustrations divide the book into six conventional sections that form the only comprehensible content: herbal, astronomical/astrological, balneological/biological, cosmological, pharmaceutical, and recipes.
Stylistic evidence points to Northern Italy or the Alpine region: a drawn castle has swallowtail (Ghibelline) merlons typical of Verona, the Sagittarius archer wears a Florentine hat, and the bathing culture mirrors balneological manuscripts popular in northern Italy at the time.
The first confirmed trace is a 1639 letter from Prague alchemist Georg Baresch to Athanasius Kircher, calling the book a 'riddle of the Sphinx'. After Baresch's death, it passed to Johannes Marcus Marci, who sent it to Kircher in 1665/66 with a cover letter stating that Holy Roman Emperor Rudolf II had purchased it for 600 gold ducats, believing it to be a work of Roger Bacon.
The Roger Bacon attribution is universally rejected today—the parchment postdates Bacon by a century. Also, the theory that John Dee owned it is now doubted.
The book lay hidden in the Jesuit Collegio Romano for two centuries. In 1912, Wilfrid Voynich bought it from the Jesuits of Villa Mondragone. After his death, it passed to his widow Ethel (daughter of George Boole), then to Anne Nill, then to dealer Hans P. Kraus, who donated it to Yale's Beinecke Library in 1969 (MS 408).
In 2020, palaeographer Lisa Fagin Davis demonstrated at least five distinct scribes wrote the manuscript, collaborating in a workshop. In 2024, multispectral imaging revealed three previously invisible columns of letters in the margin of the first page—two Roman, one Voynichese—in the hand of Marci, indicating a 17th-century decryption crib.
Professional cryptologists including William and Elizebeth Friedman, John Tiltman, and the NSA all failed. Each wave of claims has been debunked, often by subsequent statistical or historical scrutiny.
The text could be an extinct language, a dialect, or a language like Hebrew/Nahuatl written in a unique alphabet. Strengths: language-like macrostatistics. Weaknesses: no known language matches glyph distributions or the extreme positional constraints.
Ranges from simple substitution (refuted by statistics) to homophonic/polygraphic systems, nomenclators, and code-books. Friedman concluded it was a synthetic language — a small glossary encoded with systematic roots and endings. Greshko’s Naibbe cipher demonstrates feasibility with 15th-century tools. Weakness: no period cipher of this complexity exists, and the mechanism must be executable by hand across 240 pages by five scribes.
The text may be meaningless but language-flavored, generated mechanically (Cardan grille), through self-citation, or by human 'glyptolalia'. Gaskell & Bowern showed human volunteers could produce fake text matching many low-level metrics. Weaknesses: grille technique dates from ~1550; grille fails line-level structure and vocabulary size; human gibberish samples lack higher structural organization.
The Lazarus Project's multispectral images revealed Marci's marginal decryption attempt: two columns of Roman alphabet, one of Voynichese, on f1r. Only ten pages were scanned; further imaging could find more cribs. The scans also confirmed no palimpsest and no modern forgery.
Lisa Fagin Davis showed five scribes worked together, with hands correlating to Currier’s languages A and B. No scribe crossed over between dialects. This points to a workshop with standardized rule sets, ruling out solo doodling.
Michael Greshko demonstrated a hand‑executable system using dice, playing cards, and six substitution tables that converts Latin or Italian into ciphertext statistically mimicking Voynichese. This is the first proof that a medieval‑plausible mechanism can manufacture the text’s properties.
We used the Landini–Stolfi Interlinear EVA transcription (35,434 word tokens, 7,584 distinct types, 179,742 glyphs, 5,178 lines, 22‑character alphabet). Identical pipelines were applied to three control texts: Dante’s Divina Commedia (Italian), Chaucer’s Canterbury Tales (Middle English), and the Latin Vulgate.
We do not seek a decipherment (a reproducible reading). Instead, we answer the meta‑question: what kind of object is this text? Every theory must account for all measured phenomena.
Second‑order conditional entropy: Voynichese 2.14 bits vs. Italian 3.17, Middle English 3.27, Latin 3.29. This means Voynich glyphs are positionally handcuffed; a mechanism, not natural language, determines the sequence.
Low entropy alone does not prove meaning—mechanical generators can also produce it. However, it proves Voynichese is not letters behaving like letters.
Glyphs occupy rigid word positions: n is word‑final 98.9% of the time; h never starts or ends a word; y ends 87.6% of words; r closes 77.2%; q appears almost only word‑initially and is followed by o 100%. No natural script shows such determinism.
Top‑10 bigrams cover 43.3% of the text, vs. 17–23% in natural languages. Voynichese barely explores its combinatorial space. Together, these facts point to glyphs being visible fragments of hidden code units (like Morse code dots and dashes).
Word length variance is only 2.99 (vs. 4.3–6.1 in controls), clustering at 4–6 glyphs—exactly what you get from a fixed‑slot word generator. A simple slot‑grammar template already parses 57% of all tokens.
Zipf’s law slope: Voynichese −1.07, Italian −1.07, Chaucer −1.09, Vulgate −0.98. The manuscript obeys the macro‑physics of language while its micro‑physics are those of a machine. This happens when meaning is stored in unseen units (code‑book entries) that are packaged combinatorially.
Tall 'gallows' glyphs (p, t, f, k) are 3.8× enriched at line starts (p enriched 18×). Natural languages show no such line‑initial bias; something resets at each line opening.
Currier’s A/B dialect split is confirmed: dy‑ending words are 24.0% in B vs. 6.3% in A; daiin dominates A, chedy dominates B. The five scribes’ hands map perfectly to dialects, with interleaved bifolios—evidence of a coordinated workshop using two rule‑sets.
Mutual information analysis shows words cluster by section (e.g., shedys for balneological, daiin for herbal) with overwhelming significance (z=+40.24). The text is organized by topic, further supporting meaningful content.
The Timm & Schinner self‑citation model predicts strong local word similarity. We tested by comparing edit distance to preceding words against a shuffled baseline. The ratio of near‑copies (edit ≤ 1) was only 1.09× (Dante’s genuine text: 1.04×). The effect is real but far too weak to be the engine—similarity comes from the global slot grammar, not local copying.
Pure autocopying hoax is disfavored by this discriminating measurement. The same standard of proof disfavors Rugg’s grille and other simple mechanical hoaxes.
The text is rule‑generated meaningful content: a small inventory of concepts assembled via a fixed positional grammar. It was produced by a workshop using two dialects. The best‑fitting family is a structured homophonic / table cipher or a small code‑book 'synthetic language', as Friedman concluded. The meaning lives in units below the glyph and above the letter; the key is a lost code‑book. Hoax theories require such systematic discipline that they become indistinguishable from genuine authorship.
The manuscript resists decipherment not because it is empty, but because the key is a book, not an algorithm. Without a crib or the original glossary, the lock remains uncracked.