The hypotheses, put to the test of chance

Herbal-recipe book in an Alpine High German dialect (Germanic islands of the eastern Alps)

Claim. The Beinecke MS 408 is a herbal-recipe book written in an Alpine High German variety of the Germanic language islands of the eastern Alps (Carnia, Cadore, Venetian Prealps); the Timau/Tischlbong lexicon shows the highest density of correspondences (638 attested entries). It is not enciphered: it is written directly and read by the healer's visual protocol — image plus a reduced lexicon.

Test. Entropic signature (h2 ≈ 2.38 bits), a four-stage EVA→Old High German phonological pipeline, quantum clustering of phonemes (IBM Quantum), comparison with over ten medieval Alpine Germanic lexica, folio-by-folio iconographic analysis.

In progress. The low conditional entropy and the specialised, redundant lexical structure are compatible with an Alpine Germanic dialect, not with a full European language nor with a simple cipher. The full transposition into a modern language (Italian and English) is in preparation.

The medieval healer's visual protocol

Claim. The text is not continuous prose but a compressed system: the image of the plant plus a reduced lexicon (humoral quality, parts, method, organ, virtue), like the memory-aids of fifteenth-century Alpine herbalists.

Test. Comparison with the Gart der Gesundheit (1485), Cod. Pal. germ. 539 and Hildegard's Physica; reading the low entropy as the signature of the system, not as a mystery.

In progress. The low entropy (h2 ≈ 2 bits) fits a short, redundant lexicon aimed at an already-trained reader who recognised the plant from the illustration.

Combinatorial generator (Llullian wheels)

Claim. Words arise from combinatorial prefix-core-suffix wheels, in the spirit of Ramon Llull's art.

Test. Slot reconstruction from the data: 23 prefixes × 553 cores × 19 suffixes = 241,661 possible words, of which the manuscript uses 6,827.

In progress, confirmed as a generative mechanism: only 3% of the combinatorial space used, 88% of occurrences from reduced wheels. It explains 'how it is generated', it is not a decipherment.

Eleonora Matarrese — Germanic reading of the Carnic Alpine arc

Claim. Eleonora Matarrese proposes, on philological-botanical grounds and in full independence, a reading of the Voynich grounded in the Germanic linguistic area of the Alps (Carnia) and in her own sign→meaning map.

Test. A comparison of geographic conclusions, not of methods: ours is statistical-computational, hers is philological-textual.

External convergence. The coincidence of the area of origin, reached by distinct methods in full independence, strengthens the direction of the research. Convergence does not mean agreement on a decipherment.

Source: eleonoramatarrese.com ↗

Kondrak & Hauer — Hebrew 80% (Alberta 2018)

Claim. Hebrew alphagram; 80% of words in the dictionary.

Test. alphagram->dictionary + baseline (f1r, Taurus)

Excluded: not distinguishable from chance. 35%=35% (f1r), 41%=41% (Taurus). Optimising the key pushes the 'Hebrew' match to 80%, but random noise rises with it (78%): they saturate together. Such an '80%' is obtainable from almost any text.

Source: Univ. of Alberta / Wikipedia ↗

Cheshire — proto-Romance (2019)

Claim. A mixed Romance language.

Test. glyph frequency/position

Excluded: EVA-x 547x is too rare and the proposed cipher is not coherent over the whole text. Article retracted by the journal.

Source: Romance Studies (2019) ↗

Gibbs — Latin abbreviations (2017)

Claim. A line in abbreviated Latin.

Test. philological review

Excluded: below baseline, not confirmed by specialist review.

Source: TLS / Wikipedia ↗

Comparison with sacred texts (Greek Codex Sinaiticus, Latin Vulgate)

Claim. The manuscript could hide an enciphered Scripture: the Greek of the Codex Sinaiticus or the Latin of the Vulgate.

Test. conditional entropy + word lengths + Zipf, equal samples (13,100 words), with baseline

Excluded: the Voynich's conditional entropy (2.38) is too low against Latin (3.50) and Greek (3.42), which a simple cipher would have preserved. No correspondence beyond chance.

Substitution into a European language

Claim. 1:1 cipher into Latin, Greek, Hebrew, Cornish, Czech, German, Slovene, Italian.

Test. lexicon match + baseline + endings + lengths

Excluded: 0 real words at 6+ letters; equal to the baseline. No standard European language beats the random control.

Alphagram / fixed-rule anagram

Claim. Reordered letters yield a language.

Test. acyclicity + anagram baseline

Excluded: equal to the baseline (40%=40%). It is a bijection, it does not change the source entropy.

Polyalphabetic / cryptex (Vigenère)

Claim. A sliding Vigenère-type key.

Test. conditional entropy

Excluded: a polyalphabetic cipher raises entropy; the Voynich is low. Opposite direction.

Alberti cipher disk (polyalphabetic)

Claim. A rotating-disk polyalphabetic cipher (Alberti).

Test. conditional entropy h2 as the disk rotates

Excluded on statistical, not chronological grounds: it raises h2 (2.3→2.5 as the disk turns); the Voynich is ~1.7. Opposite direction, in any era.

Reflection (Atbash) and shift

Claim. Reversed alphabet shifted by L.

Test. entropy + endings

Excluded: invariant or entropy increases.

Internal A/B Rosetta

Claim. Language B is A re-enciphered.

Test. 1:1 A->B remapping + baseline

Excluded: no transformation; equal to identity.

Fillers / regular Cardan grille

Claim. Gaps filled; holes at fixed positions.

Test. position + periodicity + baseline

Excluded: no periodic grille; Rugg's method generates text without a message.

Verbose cipher of a European language

Claim. Several glyphs stand for one letter.

Test. BPE: entropy vs word length

Excluded: to get language-like entropy the words drop to 2-3 letters.

Verbose 'Naibbe' cipher (Cryptologia 2025)

Claim. A randomised homophonic verbose cipher on Latin reproduces the Voynich.

Test. constructive proof: signatures types/occ ~0.2, entropy h2 ~1.7-2, entanglement decay ~85%

Excluded as a decipherment: it reproduces the statistics starting from real Latin, but needs a repetitive source or a lossy map. It reproduces the 'how it is made', not the 'what it says'.

Cappelli abbreviations as a key

Claim. The signs derive from Latin scribal abbreviations (Cappelli): use them as a reading key.

Test. comparison with the repertoire + tails/markers + QUBO on the substitution

Excluded as a key: the 'gallows' (~50% of tokens) do not exist in Cappelli. Only some markers are akin. It stays a morphological atlas, never a reading key.

Cappelli abbreviations v2 (real Cappelli + 158,200 lemmas)

Claim. The text is abbreviated medieval Latin; the real Cappelli map plus the full lexicon yields thematically distinct sections.

Test. coverage + section-specificity JSD with 100 controls; lexicon from 5,518 to 158,200 lemmas

Excluded: coverage 12%, real JSD 0.6741 below the minimum baseline 0.6827, p=1.000. More lexicon does not produce section specificity.

Abbreviated cursive Byzantine Greek

Claim. Cursive Byzantine Greek with Greek and alchemical roots.

Test. transposition + root frequency + baseline

Excluded: only the markers (-ος, καί) come out by construction; medical roots 0.1% and alchemical 0% (baseline 0.8%/0.05%), below chance.

St. Peter's Square (Rome)

Claim. One rosette represents the elliptical square.

Test. dating of the monuments

Excluded: colonnade 1656-67, obelisk 1586 — after the codex (1404-1438).

St. Mark's Square (Venice)

Claim. The center of the rosettes represents St. Mark's.

Test. chronology + shape

Excluded: chronology compatible, but the ring shape does not match a rectangular square.

Round Jerusalem

Claim. The center represents the Christian holy city.

Test. cross of streets + gates

Excluded: shared schema, but the diagnostic cross of two roads is missing.

Labels = place names

Claim. The writing beside the rosettes are city names.

Test. similarity vs known toponyms + baseline

Excluded: below baseline (0.408 vs 0.406). No signal.

Zodiac labels = star names

Claim. The labels beside the signs are star names (an astronomical Rosetta stone).

Test. vs star names and lunar mansions + baseline

Excluded: below baseline (0.388 vs 0.372).

Proper names in the zodiac labels (Muses, constellations)

Claim. The labels beside the figures are proper names.

Test. label morphology + hapax rate + count of folios with ~9 labels

Excluded as onomastics: no zodiac folio has 9 labels (range 15-90). The matches are false positives; 91% hapax page to page, no name codebook.

Leonardo da Vinci

Claim. The genius behind the Voynich's vortices and diagrams.

Test. dating + style + paleography

Excluded: Leonardo born 1452; the parchment is 1404-1438 and the drawings are medieval, not Leonardesque.

Tucker & Talbert — American plants

Claim. New World flora.

Test. radiocarbon

Excluded: incompatible with the dating. The parchment is 1404-1438, before those plants reached Europe.

Source: HerbalGram / Wikipedia ↗

Newbold — micro-marks (1921)

Claim. A Roger Bacon cipher in the micro-strokes.

Test. review (Manly 1931)

Excluded: the reading was not reproducible on review (Manly, 1931).

Source: Wikipedia ↗

Generative AI (reading without a baseline)

Claim. AIs 'read' constellations and author signatures in the same glyphs.

Test. repeatability of the reading + baseline

Excluded: confabulation. Different readings for the same signs every time, zero baseline, invented authors. Confirmation bias, not decipherment.