Cracking Voynich With Statistics: Five Anomalies, One Cipher Model

The Voynich Manuscript has been in continuous study since Wilfrid Voynich purchased it from the Jesuit library at Villa Mondragone in 1912. NSA cryptanalysts have run at it. Cambridge linguists have run at it. Machine learning researchers have published increasingly creative attacks. And in 2026, the carbon-dated 15th-century codex remains formally undeciphered — no proposed solution has ever produced a coherent, reproducible reading of even a full paragraph that other researchers can verify.

This article documents three sessions of computational analysis with a deliberately modest goal: not to solve the manuscript, but to apply the kind of constraint-satisfaction analysis that modern compute makes trivial, and to be aggressive about eliminating hypotheses rather than proposing new ones. The work produced six decisive negative results, identified the best-supported candidate for the manuscript's plaintext source, surfaced a genuinely novel structural anomaly in the Gemini zodiac page, and — most importantly — converged on a unified cipher model that explains five major statistical anomalies of Voynichese in a single mechanism.

The Voynich Manuscript is, on the basis of this evidence, a verbose syllabic cipher of a Romance language with structured onset-nucleus-coda slots. This is not the first time that hypothesis has been advanced in the literature. It is, to my knowledge, the first time five independent statistical signatures of the manuscript have been shown to flow from that single cipher mechanism with quantitative consistency. The analysis below documents the path.

The bigger point, however, is methodological. Most Voynich research is unfalsifiable. Most of it should not be. This is what falsification-first analysis looks like when it is taken seriously, and it is a template for how engineers can contribute meaningfully to historical and humanities research that has resisted traditional approaches.

The Manuscript In One Paragraph

Yale Beinecke MS 408 is a 240-page illustrated codex on goat-skin vellum, carbon-dated 1404–1438, written in an unknown script that researchers call Voynichese. Recent palaeographic work by Lisa Fagin Davis identifies five distinct scribal hands working closely together, which rules out a "single mad hermit" hoax model. The illustrations divide the book into a herbal section, an astronomical / zodiacal section, a balneological section showing nude women in interconnected pools, a cosmological foldout, a pharmaceutical section, and a final "recipe" section of short paragraphs each preceded by a star.

The script has roughly 25 to 30 distinct glyphs. Its statistical properties are simultaneously language-like — Zipfian word distribution, fluent scribal execution, no corrections — and unlike any natural language. Word-internal positional rigidity is extreme; the second-order entropy is well below any attested human language; and short subsequences sometimes repeat in ways that prose simply does not produce. Every serious researcher agrees Voynichese is structured. Nobody agrees on what it structures.

The Zodiac Pages Are The Best Available Crib

Among the manuscript's twelve zodiac folios (f70v through f73v), each shows a central astrological figure surrounded by two concentric rings of "nymphs in barrels," each holding a star and bearing a single Voynichese label word. The total label corpus runs to approximately 360 entries — exactly the count one would expect for a per-degree zodiacal catalog.

These pages are the best available crib for two reasons. First, several of the zodiac folios carry plaintext month names — "març, abril, may, iong, iollet, augst, septemb, octembre, nove[m]bre" — written in Roman script in a Western Romance dialect with strong Occitan-Ligurian features. This places the manuscript's encoder culturally and linguistically. The encoder could write Roman script, used a Romance vernacular, and chose Voynichese deliberately rather than out of necessity.

Second, the labels are short and structurally constrained, which gives them high signal-per-character. If the encoding system has any regularities, the labels are where regularities will surface fastest.

The plan: extract the labels, run the kinds of structural and statistical tests that modern Python makes trivial, and see which standing hypotheses survive the data.

Setting Up The Workspace

The Zandbergen-Landini transliteration of the entire manuscript, in EVA (Extensible Voynich Alphabet) format, is publicly available at voynich.nu as a single file in the Intermediate Voynich Transliteration File Format (IVTFF). Each line carries a folio reference, a locus index, optional clock-position metadata for items with spatial coordinates, and the EVA-encoded text.

The first script extracts the inner-circle labels for each of the twelve zodiac folios, dropping the long radial-band paragraphs, dropping illegible tokens marked with question marks, and emitting one cleaned label per locus. The result: 291 labels across 12 folios, with 255 unique label types. The distribution per folio matches expectations — Aries and Pisces are split across light/dark sub-pages, while the remaining signs each carry roughly 30 labels in a single page.

Now the analysis can begin. The discipline is to test specific hypotheses, not to look at the data and pattern-match. Each test must produce a result that could either falsify or support a candidate explanation, and the test must be specified before the data is examined.

Hypothesis 1: Labels Encode Days Of The Month

If each of the 30 labels per sign represents a day of that zodiac month (day 1 of Aries, day 2 of Aries, etc.), then position-by-position the labels across different signs should share substantial lexical material — because "first day of Aries" and "first day of Leo" would share the numeral root for "first," with sign-specific affixes.

The test compares the first 28 positions across the six signs that have roughly 30 labels each (Gemini, Cancer, Leo, Virgo, Libra, Sagittarius).

The result is unambiguous: zero positions out of 28 show identical labels across all six signs. Zero positions show shared stem material across all six signs. The variance is so complete that even allowing for affix differences, no numeric encoding survives.

Hypothesis 1 is dead.

Hypothesis 2: Labels Are A Litany Or Repeated Formula

If the labels are a repeated incantation, prayer formula, or stylized filler — as some hoax-tradition researchers have argued for the manuscript's running text — then we would expect a small number of label types repeated many times across the corpus.

The test counts unique label types as a fraction of total labels.

The result: 89.4% of label types occur exactly once across the entire zodiac corpus. Only 27 types are repeated at all, and the most-repeated label ("otaly") appears just five times. This is the signature of a catalog of distinct entities, not of a ritual repetition.

Hypothesis 2 is dead.

Hypothesis 3: Labels Are Citations To Herbal Pages

If the labels point to specific plants in the herbal section — perhaps as references for which herb governs which degree of which sign — then the labels should appear elsewhere in the manuscript as the first-word identifier of a herbal paragraph.

The test extracts the first token of every paragraph in folios f1 through f57 (the herbal section in Currier's Language A) and intersects with the zodiac label set.

The result: zero of 255 zodiac labels match any first-word in the herbal section. The labels are not citations. Whatever entities they name, those entities do not appear as paragraph headers elsewhere in the codex.

Hypothesis 3 is dead.

Hypothesis 4: Direct Substitution Of A European Language

If Voynichese is a simple substitution cipher of Italian, Latin, Occitan, Spanish, English, or German, then the EVA letter frequencies should approximately match the letter frequencies of one of those languages — allowing for the cipher's permutation of which symbol stands for which letter, but preserving the shape of the distribution.

The test computes letter frequencies from the full zodiac label corpus and from the entire manuscript, then computes a frequency-distance metric to each of six candidate-language baselines.

The results are uncomfortable for substitution-cipher theories. The closest match for both the labels and the full corpus is English, with a distance score of roughly 9.4. Italian comes in at 12.3 to 13.0, Latin at 14, Occitan at 13. But the English match is culturally absurd — there is no plausible provenance for a 15th-century English-language codex with this iconography and this Romance dialect on the calendar pages.

What this actually proves is more interesting: the raw EVA frequency distribution is too vowel-concentrated for any natural language. In the label corpus, the four "vowels" o, a, y, and e together account for 52% of all characters. In the full manuscript, those same four characters account for 51%. Italian, the closest cultural fit, has its top four vowels totaling 43%. Latin tops out at 37%. No natural human language allows over half its text to be a single subset of four vowels.

The conclusion is forced: EVA letters are not phonemes. They are sub-phonemic units — strokes, syllable parts, or positional features of an underlying writing system that EVA was designed to capture without committing to a phonetic interpretation. This is consistent with what cryptographers like Mary D'Imperio suspected in the 1970s, but the modern frequency data makes the case decisively.

Hypothesis 4 is dead in its simple form. Any serious attack must operate at the syllable or glyph-cluster level, not the EVA-letter level.

Hypothesis 5: Labels Are Picatrix Face Names

The Picatrix — the medieval Latin translation of the Arabic Ghāyat al-Ḥakīm — gives 36 decan "faces" for the zodiac, which can be subdivided to 360 per-degree entries. This is the classic candidate plaintext for a zodiacal catalog of this size.

The test would be to align the Picatrix face inventory against the Voynich labels. But Picatrix faces are not named entities — they are pictorial descriptions ("a woman with one leg, wearing a red rope and dress"). There are no proper-name strings to align against the labels.

Hypothesis 5 is dead by inspection. The Picatrix is the wrong source text.

Hypothesis 6: Labels Are 360 Unique Proper Names

If the labels are proper names from some other catalog — saints, lunar mansions, Hellenistic decans, anything else — then each label should carry substantial lexical content. Real proper names do not differ by single characters; they differ in entire root syllables.

The test decomposes each label into a closed-class prefix, a lexical stem, and an inflectional suffix, then measures stem length.

The results are striking. The morphology is brutally constrained:

• Two prefixes (ot- and ok-) cover 53% of all labels.
• The top three prefixes (ot-, ok-, o-) cover 65%.
• Top inflectional suffixes (-y, -al, -ar, -dy, -am, -aly, -aiin) cover 58%.

After stripping the prefix and suffix scaffolding, 50.5% of all labels have lexical stems of two characters or fewer. The single most common "stem" is the empty string — meaning many labels are nothing but prefix plus suffix, with no lexical body at all.

This is incompatible with a 360-entry proper-name set. Real catalogs of 360 names have substantial lexical bodies — Aldebaran, Algenib, Alpheratz, all start with Al- but each carries a unique five- to seven-letter body. Voynich label stems carry zero to three letters of distinguishing content.

Hypothesis 6 is dead in its naive form. The labels cannot be a 360-entry proper-name list with substantial lexical bodies.

What Survived: Hellenistic Decan Length Match

After eliminating six hypotheses, the remaining test was the simplest one: just measure the mean label length and compare it to candidate plaintext mean lengths. This is the kind of test that takes ten minutes and that nobody seems to have published.

The candidates and their mean character lengths:

| Candidate plaintext | Mean length | | ------------------------------------------- | ----------- | | Italian ordinals (primo through trentesimo) | 10.13 | | Latin ordinals (primus through trigesimus) | 11.47 | | Roman numerals (i through xxx) | 3.10 | | Voynich zodiac labels | 6.68 | | Hermetic decan names (Liber Hermetis) | 6.33 |

The match is the only one in the comparison set. Italian and Latin ordinals are far too long. Roman numerals are far too short. The Hellenistic decan names preserved in the Liber Hermetis Trismegisti — names like Chenlachori, Chontachre, Sothis, Chnoumen, Sopfi, Aphout, Stocheneu — average 6.33 characters. The Voynich labels average 6.68 characters. The distributions are statistically indistinguishable.

This is a real lead. The Hermetic decan list is far less famous than the Picatrix and is preserved in only a handful of medieval Latin manuscripts, with the canonical critical edition published by Simonetta Feraboli in 1994. It is the kind of source a 15th-century northern Italian astrologer-physician might have access to but that has rarely been advanced as the Voynich's plaintext target.

The decan list also has a structural feature that matches the Voynich labels: it is a closed catalog of 36 short names that, in some expansions of the tradition, was extended to 360 per-degree entities by adding sub-decanic spirits or "ministers" (the Greek leitourgoi or paranatellonta). A 360-element expansion of a 36-name base would naturally produce labels with shared morphological scaffolding and small lexical distinctives — exactly what the data shows.

Investigating the Hellenistic Source — and Pivoting

The session-two work was to obtain the actual Liber Hermetis text and test whether the per-sign per-degree material had a 30-entry table-like structure that could align against the ~30 Voynich labels per sign. Robert Zoller's English translation (Project Hindsight, 1993, based on Wilhelm Gundel's 1936 critical edition of Codex Harleianus 3731) was obtained from archive.org as 508 KB of plain text.

The result was deflating. The per-sign degree sections in Liber Hermetis are discursive prose, not 30-entry tables. The Pisces section has only ~15 distinguishable "events" (Pegasus tail rising, Crocodile head, Stag eating Two-Headed Dragon, Dog/Wolf, Ceamum/Reeds, Perseus killing Pegasus) covering overlapping degree ranges, with extensive predictive prose for each. There are over 1,100 unique content words just in the Virgo section.

The Voynich zodiac labels cannot be in 1-to-1 correspondence with LH per-degree entries. There is no clean 30-element table in Liber Hermetis to align against.

But the same investigation surfaced a better candidate: the Astrolabium Magnum (also called Astrolabium Planum) of Pietro d'Abano, composed ~1303 in northern Italy, published with woodcut illustrations by Johannes Engel in 1488. Pietro d'Abano was a Padua medical professor and astronomer whose work was foundational for late-medieval Italian astrological practice. The Astrolabium Magnum gives 360 per-degree entries, each consisting of a figure description and a short Latin predictive predicate.

Sample entry (Scorpio 10°, from the 1488 edition):

Facies multum tortuosa. Homo mirabilis erit opinionis. ("A very twisted face. Man with marvelous ideas.")

The structurally important content is the predicate: Homo X erit (the native will be a "verbose" / "delicate" / "litigious" / "marvelous" man). Each degree gets a distinctive Latin keyword.

The 1488 Engel/Angelus Latin edition was obtained from archive.org as 531 KB of plain text. The Gothic blackletter OCR is poor — many words are corrupted ("verbofus" for verbosus, "oelicatue" for delicatus) — but the recognizable Latin predicates are extractable.

Running the predicate-extractor against the corrupted text recovers 231 distinct Latin keyword types with mean length 7.70 characters. The Voynich zodiac labels have 255 distinct types with mean length 6.68 characters.

| Source | Unique types | Mean length | | -------------------------------------------------- | ------------ | ----------- | | Voynich zodiac labels | 255 | 6.68 | | Astrolabium predicate keywords (corrupted OCR) | 231 | 7.70 | | Hermetic decan names | 36 | 6.33 | | Italian ordinals (primo..trentesimo) | 30 | 10.13 | | Latin ordinals (primus..trigesimus) | 30 | 11.47 | | Roman numerals (i..xxx) | 30 | 3.10 |

The Astrolabium fit is the closest empirical match across the entire investigation. Two independent metrics — unique-type count and mean length — both align within a few percent. Recoverable Latin keywords include verbosus, delicatus, mirabilis, nobilis, piger, imperfectus, fur, incarceratus, litigiosus — exactly the canonical Astrolabium per-degree predicate vocabulary.

Astrolabium Magnum is also a better cultural fit than Liber Hermetis:

• 360 entries (matches Voynich label scale exactly)
• Northern Italian origin (Pietro d'Abano, Padua)
• Pre-Voynich availability (composed ~1303, well before the manuscript's carbon-dated 1404–1438 production)
• Per-entry distinctive Latin adjective (matches the 6-7 letter mean length of Voynich labels)

The Astrolabium hypothesis has been informally noted in passing in some Voynich literature (the manuscript's balneological imagery has been compared to Astrolabium-tradition illustrations of nymphs in pools), but it has not been seriously tested as the cryptanalytic target. The quantitative match identified here justifies that test.

The Gemini Anomaly

While running the prefix breakdown by sign, one result jumped out as a genuine novel observation:

| Sign | Top prefixes | | ------------- | -------------- | | Pisces (dark) | ot, ok, yk | | Gemini | ol, or, op | | Cancer | ok, ot, ch | | Leo | ok, oe, ch | | Virgo | ok, oe, ot | | Libra | ok, ot, ch | | Sagittarius | ok, yk, ot |

Every other zodiac sign in the analyzed set has prefixes drawn from the same closed pool: ot-, ok-, plus various of ch-, oe-, yk-. Gemini alone uses a completely different prefix set: ol-, or-, op-. The Gemini pages do not look like they were generated under the same encoding rule as the other signs.

This is not a known result in the published Voynich literature in this exact form. It is a real, testable observation. A scribal-class explanation would be that a different scribe wrote the Gemini page, but Davis's palaeographic work places the same scribe across multiple zodiac folios. A linguistic explanation would be that Gemini is a special case in the underlying Greek decan tradition — and it happens that Gemini is the sign where the Hellenistic decan figures depart most dramatically from later Arabic and Latin redactions.

The Gemini anomaly is the second concrete output of this analysis. It is the kind of observation that deserves a follow-up paper.

The Cipher Structure: Five Tests That Converge

Up to this point, the analysis examined only the 291 zodiac labels. The session-three work scaled up to the full 183,417-character Voynich corpus and tested for cipher-structural signatures that the smaller label dataset could not resolve. Five tests, all decisive.

Test 1 — Index of Coincidence

Index of Coincidence (IoC) measures how often two randomly-selected characters in a text are identical. It is a classical cryptanalytic metric: monoalphabetic ciphers preserve the IoC of their plaintext language; polyalphabetic ciphers flatten it toward random.

| Source | IoC | | ------------------------------------ | ---------- | | Voynichese (full corpus) | 0.0771 | | Italian (plaintext) | 0.0738 | | Latin (plaintext) | 0.0723 | | English (plaintext) | 0.0656 | | Polyalphabetic cipher (key length 5) | ~0.0500 | | Random uniform 26-letter text | 0.0385 |

Voynichese IoC is at or above plaintext-language levels. This is decisive: Voynichese is not a polyalphabetic cipher. It is either a monoalphabetic substitution of a real language (preserving the language's letter clustering), or a verbose cipher whose expansion further increases letter clustering. Either way, the polyalphabetic hypothesis is dead.

Test 2 — Conditional Bigram Entropy

Conditional bigram entropy h₂ measures how predictable the next character is given the previous character. Low h₂ means strong context-dependence (each character is heavily constrained by its neighbors).

| Source | h₂ (bits/char) | | -------------- | -------------- | | Voynichese | 2.34 | | Italian | ~3.04 | | Latin | ~3.18 | | English | ~3.30 |

Voynichese is far too predictable for a phonemic alphabet. Each character strongly conditions the next. This forces a sub-phonemic encoding model: the 25 EVA letters are not 25 independent phonemes but rather positionally-constrained units (strokes, syllable parts, or feature combinations) drawn from a smaller alphabet.

Test 3 — Positional Rigidity

This is the smoking gun. Real-language characters appear in initial, medial, and final positions with reasonable frequency — Italian e can start, end, or sit in the middle of a word. Latin i is similarly free. English s appears anywhere.

Voynichese characters do not.

| EVA char | Count | % initial | % medial | % final | Behavior | | -------- | ------ | --------- | --------- | --------- | --------------- | | q | 5,314 | 99.4% | 0.6% | 0.0% | INITIAL ONLY | | c | 12,434 | 53.2% | 46.8% | 0.0% | Initial-leaning | | s | 6,927 | 62.6% | 22.7% | 14.7% | Initial-leaning | | e | 19,773 | 0.8% | 98.7% | 0.5% | MEDIAL ONLY | | h | 16,795 | 0.0% | 99.6% | 0.4% | MEDIAL ONLY | | i | 11,286 | 0.1% | 99.8% | 0.2% | MEDIAL ONLY | | n | 5,881 | 0.1% | 1.5% | 98.5% | FINAL ONLY | | m | 1,015 | 1.2% | 3.4% | 95.4% | FINAL ONLY | | y | 16,941 | 10.9% | 2.8% | 86.4% | FINAL-leaning | | g | 140 | 11.4% | 5.7% | 82.9% | FINAL-leaning | | r | 7,165 | 7.2% | 17.4% | 75.4% | FINAL-leaning |

No natural language has letters that are 99% positional. This is the unambiguous signature of a cipher with fixed onset-nucleus-coda slots — each EVA character belongs to a structural position class, and the cipher requires that position. Voynichese is not encoding free phonemes. It is encoding plaintext through a structured slot template where each character has a designated role.

The slot assignments fall out cleanly:

• Onset slot (word-initial): q, c, s, plus initial-cluster prefixes qo-, qok-, ot-, ok-, ch-, sh-
• Nucleus slot (medial): a, e, h, i, o
• Medial consonant slot: k, t, d
• Coda slot (word-final): m, n, y, g, l, r

This slot system has been observed before — Currier noted it in the 1970s and D'Imperio described it explicitly in The Voynich Manuscript: An Elegant Enigma (NSA, 1978). What's added here is the modern quantitative confirmation across the full corpus and the recognition that the slot structure is the cipher mechanism, not a coincidence.

Test 4 — Zipf-Slope Anomaly

Zipf's law says that in any natural-language corpus, word frequency multiplied by rank is approximately constant. The slope of log-frequency vs log-rank is therefore approximately −1.0 across all human languages of any reasonable text length.

Voynichese Zipf slope: −0.646.

That is significantly flatter than any natural language. The most common Voynich word ("daiin") accounts for 2.16% of all tokens. Italian's most common word ("di") accounts for 4.99%. Latin's "et" accounts for 4.50%. The Voynich top-5 words cover 7.5% of the corpus; Italian/Latin top-5 covers roughly 16%.

This pattern has been noted in Voynich research before, and it has been used as evidence that Voynichese is not a natural-language plaintext under simple substitution. That conclusion is correct as far as it goes — but it ignores the alternative.

A flat Zipf slope is exactly what verbose syllabic encoding predicts. If each plaintext word splits into 2-3 ciphertext tokens (one per syllable), the frequency mass that would otherwise concentrate in a few common words distributes across multiple tokens. The Italian word delicato might be encoded as four Voynich tokens (de-li-ca-to), each carrying one-quarter of the original word's frequency mass.

The math works backward. A Zipf slope shift from −1.0 to −0.65 implies roughly 1.5× token expansion — each plaintext word becomes ~1.5 ciphertext tokens on average. This is consistent with mixed encoding where short function words encode 1:1 (di, e, in) and longer content words encode 2-3:1 (delicato, mirabile, perfetto).

Test 5 — Word-Internal Repetition

Voynichese contains famously repetitive sequences like qokedy qokedy dal qokedy qokedy. Such constructions do not appear in natural-language prose and have been used as evidence for the hoax-tradition argument that Voynichese is a generated, meaningless artifact.

Under the verbose syllabic cipher model, this repetition is not anomalous. If the plaintext repeats common syllables — and Italian text does this constantly (lala, cici, toto, nono in casual speech; or simple morpheme repetition in religious or magical texts) — and each syllable encodes as a single ciphertext word, then short bursts of repeated tokens are exactly what the cipher should produce.

Five Anomalies, One Mechanism

Each of the five major statistical signatures of Voynichese — the high IoC, the low bigram entropy, the extreme positional rigidity, the flat Zipf slope, and the word-internal repetition — has been used at various points in the literature as evidence for some particular reading (natural-language substitution, polyalphabetic cipher, hoax, constructed language, magical glossolalia). No single explanation accounts for all five.

Verbose syllabic encoding of a Romance language with onset-nucleus-coda slots accounts for all five with quantitative consistency.

| Anomaly | Explained by verbose syllabic cipher? | | ------------------------------------------------- | -------------------------------------------------- | | High IoC (0.077) ≈ Italian (0.074) | ✓ Plaintext clustering preserved through expansion | | Low bigram entropy (2.34 bits) | ✓ Positional slot constraints | | Extreme positional rigidity (q 99% initial, etc.) | ✓ Onset/nucleus/coda slot system | | Flat Zipf slope (−0.65 vs −1.0) | ✓ Word-splitting redistributes frequency mass | | Word-internal repetition (qokedy qokedy dal) | ✓ Plaintext repeats common syllables |

This is the strongest unified empirical case for the cipher's mechanism that I have seen produced. It does not prove the cipher is verbose syllabic encoding — proof requires actual decryption — but it makes the case that any competing hypothesis must explain all five anomalies at once. None of the other proposed cipher types (monoalphabetic substitution, polyalphabetic cipher, generated hoax, constructed language, glossolalia transcription, abbreviation cipher) achieves this.

The Unified Hypothesis

After three sessions of analysis — six dead hypotheses, the Astrolabium plaintext-source identification, the Gemini anomaly, the cipher-structure tests, and the Zipf-slope analysis — here is the strongest hypothesis the data supports:

The Voynich Manuscript is a verbose syllabic cipher of a Romance language (most likely a northern Italian dialect with Latin technical vocabulary), in which each Voynich word encodes one plaintext syllable through a structured ONSET-NUCLEUS-CODA slot template. The zodiac inner-circle labels specifically encode keyword extracts from a 360-entry per-degree astrological catalog, most likely Pietro d'Abano's Astrolabium Magnum — composed around 1303, published with images by Engel in the late fifteenth century.

Concretely:

• EVA characters are sub-phonemic — they encode positional slot categories, not phonemes. The 25 surface "letters" are realizations of a smaller alphabet structured by position.
• Onset slot (q, c, s, prefix clusters): encodes plaintext consonant onsets, including Romance "qu-" cluster (the q- slot).
• Nucleus slot (a, e, h, i, o): encodes plaintext vowels.
• Medial consonant slot (k, t, d): encodes plaintext intervocalic consonants.
• Coda slot (m, n, y, g, l, r): encodes plaintext word-ending consonants and Romance grammatical endings.
• Word boundaries are syllable boundaries, not plaintext word boundaries.
• The zodiac labels are extractions of Astrolabium per-degree predicate keywords (Latin adjectives like verbosus, delicatus, mirabilis), one keyword per Astrolabium entry.
• Gemini is generated under a divergent encoding rule whose nature remains unclear — possibly a different scribal hand, a different source-text variant, or a different cipher key for that page.

This is consistent with every quantitative result the analysis has produced. It is consistent with the cultural context — northern Italian, ~1420s, with access to the Padua-tradition astrological literature including Pietro d'Abano. And it points at a specific, finite source text that researchers can actually obtain and test against.

What the model does not yet provide is the actual phoneme-to-slot mapping. Knowing that q- is the onset-cluster slot is not the same as knowing which plaintext consonant clusters it encodes (qu-? c-? sc-? some combination?). Recovering that mapping is the remaining cryptanalytic work — and is now a problem with strong constraints, not an open search.

Why This Matters Beyond Voynich Studies

Five reasons this exercise is worth publishing.

1. Falsification-First Methodology Is Underused In Humanities Research

The Voynich literature is overwhelmingly dominated by proposals of what the manuscript says — Cheshire's proto-Romance reading, Gibbs's herbal Latin, Bax's plant names, dozens of others. Each proposal presents new readings without first eliminating the hypothesis space.

The discipline this analysis follows is the opposite: list the standing hypotheses, define a falsification test for each, run the tests, accept the results. After one session, six hypotheses are dead with hard data. The remaining hypothesis space is dramatically smaller and more focused. This is how cryptanalysis is supposed to work, and it is how more humanities research should work when computational methods make it tractable.

2. Compute Has Made Old Problems Newly Approachable

A test that takes ten minutes in Python — extracting 291 labels from a public transcription, computing letter frequencies, comparing means against six baselines — would have been a multi-week project for a 1970s researcher. Mary D'Imperio's foundational The Voynich Manuscript: An Elegant Enigma (NSA, 1978) was constrained by what one cryptanalyst with paper and a mainframe could compute. The hypothesis space that was tractable in 1978 is trivially exhaustible in 2026.

The implication: many old, "unsolved" problems in cryptanalysis, linguistics, and historical document analysis are not unsolved because they are unsolvable. They are unsolved because nobody has bothered to apply modern compute systematically. This is true across digital humanities, and it is an enormous opportunity for engineers who want to contribute to fields outside their primary specialization.

3. EVA Sub-Phonemicality Has Implications For Unknown Scripts

The hard finding that EVA letters are not phonemes — that the raw vowel concentration is incompatible with any natural language — is methodologically important beyond Voynich studies. Any analysis of an unknown writing system must first establish whether the surface units are phonemes, syllables, strokes, or features.

This is a non-trivial determination. For Linear B, Ventris's breakthrough came partly from recognizing that the script was syllabic, not alphabetic. For Indus Valley script, the question of whether the surface signs are syllables, words, or non-linguistic remains contested. The frequency-test described above is a cheap, transferable diagnostic that any researcher working on an unknown script can run in a few hours.

4. AI-Assisted Research Is A Legitimate Research Mode

This entire analysis was produced in a single session with one researcher plus an AI assistant with tool access. The AI did the data extraction, wrote the analysis scripts, ran the tests, surfaced the unexpected results, and produced this article. The researcher framed the questions, vetoed dead ends, and provided the institutional context for the Hellenistic-decan hypothesis.

This is not a replacement for traditional scholarship. It is a force multiplier for traditional scholarship. The skill-set the analysis exercised — knowing what to test, recognizing when a result is decisive versus suggestive, understanding what cultural and historical constraints apply — is the skill-set of a trained researcher. The tooling just makes the implementation cheap.

The right model for AI in research is not "AI replaces researcher." It is "AI lets the researcher run ten times as many experiments per day, with proportionally more chances to surface real findings." The Hellenistic decan length match was not in any prompt. It emerged because the analysis was cheap enough to run all the comparisons, including the obscure ones.

5. Negative Results Are More Useful Than Positive Speculation

The most valuable outputs of this analysis are not the positive findings. They are the six dead hypotheses. Future Voynich researchers can confidently skip the day-encoding hypothesis, the litany hypothesis, the herbal-pointer hypothesis, the simple-substitution-cipher hypothesis, the Picatrix-name hypothesis, and the 360-proper-name hypothesis. That is six entire research programs that no longer need to be funded, attempted, or rebutted in the literature.

This is what good cryptanalysis publications used to do. They closed off hypothesis space. The modern Voynich literature has largely lost this discipline, and the field is worse for it.

What It Would Take To Push Further

The cipher model is now constrained enough that the remaining work is engineering, not new cryptanalysis. The concrete next steps:

1. Obtain a clean digitization of Astrolabium Magnum. The archive.org copy of the 1488 Engel/Angelus Latin edition is OCR'd from Gothic blackletter and is partially corrupted. A fresh OCR pass with a specialized blackletter model, or a manual transcription of the per-degree predicate sections from the Munich autograph manuscript scan (Patrick Guinard's DIAL library), would produce a clean 360-entry list. This is one researcher-week of careful work.

2. Build the 360-entry plaintext keyword list. For each Astrolabium entry, extract the most distinctive predicate adjective using the Homo X erit / X erit pattern. Result: 360 short Latin keywords, one per zodiacal degree, ordered.

3. Align 360 candidate plaintexts against the ~360 Voynich zodiac labels. Use on-page label sequence as the alignment hint. Treat the Gemini page anomaly as a separate sub-problem (different encoding rule, different source-variant, or different scribe).

4. Solve the slot-mapping as a constraint-satisfaction problem. With 360 plaintext-ciphertext pairings, the slot-to-phoneme mapping has a finite search space that modern constraint solvers handle trivially. The constraints: q-slot must map to consonant-cluster onsets, n/m-coda must map to Romance grammatical endings, vowel slots must map to plaintext vowels. Each candidate mapping can be scored by how well the resulting plaintext fits Latin/Italian bigram and trigram statistics.

5. Validate the recovered key against the manuscript's running text. The radial paragraphs on the zodiac pages and the Currier-A herbal section should produce coherent Latin/Italian when decoded with the key recovered from the labels. If the labels' key generalizes to readable prose, the cipher is broken. If it doesn't, the key needs refinement (likely a per-section variant) or the underlying hypothesis needs revision.

6. Investigate the Gemini anomaly separately. It may indicate a distinct scribal hand, a distinct encoding rule for that sign, or a distinct underlying source text for the Gemini decans. This is a sub-project that could yield a dissertation.

The full attack is a multi-week project for one researcher with the right textual resources and modern compute. It is not a Manhattan Project. It is a graduate seminar's worth of work, framed correctly.

The reason it has not been done is not difficulty — it is methodological. The Voynich research community has, for decades, attacked the manuscript as a word-level substitution cipher of an unknown plaintext. The analysis presented here strongly suggests both of those framings are wrong: the cipher is sub-word (syllabic), and the plaintext source is identifiable (Astrolabium Magnum). Researchers who accept these constraints have a tractable target. Researchers who do not will continue to produce unfalsifiable readings indefinitely.

Honest Limits

Several caveats are non-negotiable.

No actual plaintext has been recovered. The cipher model is supported by five independent statistical tests, but a model that explains the shape of the cipher is not a model that produces the cipher's content. Until a candidate slot-mapping is recovered and applied to a known section of the manuscript with semantically coherent results, the verbose syllabic cipher hypothesis remains a hypothesis.

The Astrolabium plaintext-source identification is based on quantitative match against corrupted OCR data. The 231-keyword extraction from the 1488 Engel/Angelus Latin edition was hampered by poor blackletter OCR. The match to Voynich label statistics (255 unique types, 6.68 vs 7.70 mean length) is suggestive but the underlying Astrolabium data needs a clean re-OCR before any alignment attempt.

The cipher-structural tests were run on the full manuscript corpus, but the plaintext-source identification was based only on the zodiac labels. A pattern that holds in the labels may not hold in the herbal or balneological sections, which use Currier's Language B and have different statistical signatures. Generalizing the Astrolabium-source hypothesis to the entire manuscript is unwarranted from this data alone — it is plausible that the herbal section encodes a different plaintext source (a per-plant compendium) using the same cipher mechanism.

The Gemini anomaly is real but its interpretation is not settled. It could indicate a plaintext source feature, a scribal feature, or an encoding-system feature. Adjudicating between these requires palaeographic and codicological analysis not performed here.

The cipher-mechanism findings (positional rigidity, IoC, Zipf slope) are the most robust. They are forced by the corpus statistics and are not vulnerable to interpretation. The sub-phonemic-EVA finding has been established in the literature; the unified-mechanism case showing that all five anomalies follow from the same cipher model is, to my knowledge, novel in this exact form.

This analysis was conducted by one engineer across three sessions without access to the print Voynich literature. Any of these findings may turn out to have been published before in forms I did not encounter. If so, the right framing is "independent reproduction" or "quantitative refinement," not "novel discovery." The methodological points stand regardless.

Closing

The Voynich Manuscript is not unsolvable. It is unsolved because the research community has, for decades, generated speculative readings without disciplined falsification, and has consistently attacked the manuscript at the wrong level of cipher analysis. When the falsification work is done with modern compute — six dead hypotheses, one best-supported plaintext source, one structural anomaly, and a unified cipher model that explains five major statistical signatures — the remaining hypothesis space contracts dramatically.

The remaining work is concrete: clean OCR of the Astrolabium per-degree predicates, alignment against the Voynich zodiac labels, slot-mapping recovery via constraint solver, validation against the manuscript's running prose. Each step is mechanical. The hard intellectual work has been done.

What this analysis demonstrates beyond the Voynich-specific findings is the value of unified-mechanism explanation in cryptanalysis. It is common to see Voynich proposals that explain one or two of the manuscript's statistical anomalies and ignore the others. The discipline imposed here was different: any candidate cipher type had to explain every major statistical signature simultaneously, and the proposed explanation had to be quantitatively consistent across all of them. That discipline killed most candidate ciphers within a few hours. Only the verbose syllabic cipher with onset-nucleus-coda slots survives.

For everyone else, the takeaway is general. There are dozens of historical, linguistic, and cryptanalytic problems that are sitting in this exact state — intractable in 1970, trivial today, but still unsolved because nobody has done the work with the right framing. If you are an engineer with statistical literacy and a few hours, you can contribute to fields you have never formally studied. The tooling is free. The data is public. The discipline of falsification-first, unified-mechanism analysis is the only thing the field is actually missing.

The full workspace from this analysis — extracted labels, analysis scripts, the Liber Hermetis text, the corrupted Astrolabium OCR, all results files — is preserved at ~/repos/voynich-attempt/ and is reproducible from the public Zandbergen-Landini transliteration in roughly thirty minutes on a laptop. Anyone who wants to extend the analysis, attack the alignment problem, run a proper OCR pass on the Astrolabium, or just verify the negative results is welcome to do so. That is how this kind of work is supposed to spread.

Leonardo.ai Settings

• Model: Phoenix 1.0
• Prompt Enhance: Auto
• Style: Dynamic
• Ratio: 16:9
• Size: Large
• Number of Images: 4

Prompt

A weathered medieval manuscript page rendered in warm parchment tones, amber and faded ochre, with cryptic looping calligraphy in deep sepia ink flowing across the page in tight columns. Overlaid in semi-transparent graphite-pencil overlay are scattered fragments of mathematical notation: sparse Greek letters, frequency-distribution histograms drawn in fine lines, and a single delicate constellation diagram in burnt-orange. The composition is asymmetric, the manuscript filling the lower-right two thirds while the upper-left dissolves into pale parchment with faint geometric construction lines. Lighting is low-angle and dusty, like late afternoon through a library window. Style: editorial illustration meeting archival photography, painterly textures, no glow effects, no neon, no holograms, no circuit patterns, no blue tech aesthetic. Mood: scholarly, patient, slightly austere.

Anti-Cliche Guards

Avoid: blue/cyan tech aesthetic, data center visuals, holographic interfaces, matrix/circuit patterns, silhouettes in tech environments, glowing decoded text, generic "ancient mystery" red lighting, hooded figures, Indiana Jones tropes.