A 600-Year-Old Book No One Can Read: Statisticians Finally Prove It Really Contains "Two Languages"
> Paper: A Quantitative Confirmation of the Currier Language Distinction > Author: Christophe Parisel > arXiv: 2604.25979 [cs.CR, cs.CL] > Published: 2026-04-28
1. A Book from a Basement
In 1912, Polish-American book dealer Wilfrid Voynich entered the basement of a Jesuit monastery in northern Italy and pulled a small book from a dusty chest—about 240 pages, 23.5 cm long and 16.2 cm wide, bound in vellum.
He opened the first page and froze.
The book was filled with unfamiliar symbols: curving letters, plant illustrations that seemed alien, naked human bodies soaking in strange tubes, precise astronomical dials, and a "script" no one could read. Carbon-14 dating later showed the vellum was made between 1404 and 1438—late medieval Europe, decades before Leonardo da Vinci was born.
But what drove cryptographers, linguists, and historians crazy wasn't the bizarre illustrations. It was the text itself.
For over six centuries, countless people tried to decipher it. The head of British WWII codebreaking, a member of the Enigma-cracking team, tried. An entire department of the NSA studied it. In the computer age, machine learning algorithms were fed these symbols. Everyone failed.
This is the Voynich Manuscript—widely recognized as the world's most mysterious manuscript.
In April 2026, an independent researcher named Christophe Parisel published a paper using cold, hard statistics to prove something remarkable: the manuscript's "text" really does split into two distinct systems. Not a guess, not an impression—a mathematical necessity.
2. Currier's Intuition: A Cryptographer's Keen Nose
The story begins in 1976.
Prescott Currier was a US Navy cryptographer who deciphered enemy codes during WWII and later worked at the NSA, with a bloodhound's sensitivity to statistical patterns. At a 1976 academic symposium on the Voynich Manuscript, he stood up and said something that changed the field:
"This book contains more than one 'language.'"
Based on his own statistical counts, he divided the manuscript into two classes—later called Currier A and Currier B. His evidence wasn't subjective "feelings" but concrete numbers:
- On A-type pages, the symbol combinations "chol" and "chor" appear with very high frequency, often repeated consecutively; they are nearly absent in B.
- On B-type pages, the word-final "dy" is rampant; in A it barely exists.
- "chaiin" is rare in B, moderately frequent in A.
- The most common word in A is "daiin"; in B it is "chedy"—which never appears in A.
- Split-Markov models (assuming A and B are each internally uniform) systematically underestimate per-folio variance by factors of 1.5 to 6. This means A and B each contain substructure or gradients.
- Spatial asymmetry: forward spatial splits (predicting later folios from earlier) achieve 79.6% accuracy, but reverse splits only 41.3%. Not signal failure—the manuscript's temporal structure is itself asymmetric: early parts are mostly A, later parts mostly B. This suggests a possible composition order.
- Many "intermediate forms" exist; A and B are not two isolated islands but the two ends of a spectrum.
- Why some character pairs switch near-categorically between A and B while others are completely unaffected?
- Why "chedy" is nearly absent in A while "daiin" still occurs in B?
- Why a single quire can contain both A and B folios?
Currier immediately clarified: his "languages" did not necessarily mean different natural languages like Greek vs. Latin. He meant statistical feature patterns—like two people's handwriting: even in the same language, letter slant, spacing, and joining habits differ sharply.
More strikingly, Currier noticed the manuscript's handwriting itself split into two classes: A's letters are large and rounded, B's small and compact. Later paleographer Lisa Fagin Davis subdivided the scribal hands into five—but remarkably, no single hand wrote both A and B. Scribes 1 and 4 wrote A; Scribes 2, 3, and 5 wrote B.
Yet this discovery was like a key thrown into the ocean—everyone knew it might matter, but for fifty years almost no one rigorously verified it with modern quantitative methods.
3. Why Was Verification So Hard?
Parisel's paper identified three core challenges:
First, natural fluctuation? Local statistics of any long text fluctuate. Could the A/B difference just be random noise—"the first half of the book happens to differ from the second"?
Second, confounding by physical structure? Currier's original classification correlates highly with the manuscript's quire boundaries. The A/B distinction might be a byproduct of "different quires written by different scribes," not an intrinsic property of the text.
Third—and hardest: if Currier had never existed, could a completely uninformed algorithm discover the distinction on its own? Would an unbiased, data-driven analysis independently derive the same A/B split?
Parisel answered all three questions in a 16-page paper.
4. The Magic of the Beta-Binomial Mixture Model
Parisel's core weapon is the Beta-Binomial mixture model—a generative probabilistic model commonly used for overdispersed count data.
In plain language: if the Voynich Manuscript uses some substitution cipher system, and A and B represent two different substitution rules, then visually similar character pairs should show systematic frequency-ratio differences between A and B.
For example: suppose characters "d" and "l" are two variants (allographs) of the same cipher letter. If "d" dominates in A and "l" dominates in B, then the d/l ratio should jump at the A/B boundary.
Parisel computed substitution ratios for multiple character pairs across 185 folios, then did three things:
First: shuffled the labels. He randomly permuted Currier's A/B labels a thousand times. Random labels almost never reproduced Currier's observed pattern (p < 0.001). The difference is not an illusion.
Second: single-language Markov simulation. Assuming the whole book came from one uniform statistical source, he generated 200 "fake manuscripts" by computer. None replicated the real manuscript's A/B difference pattern. The book cannot have been written "by one person, one rule, from start to finish."
Third: unsupervised clustering. This is the paper's most stunning part. Parisel hid Currier's A/B labels entirely and fed only raw character counts to the Beta-Binomial mixture model, asking it to decide—without any prior classification—how many statistical regimes the book contains.
The Bayesian Information Criterion (BIC) clearly selected k = 2—two regimes. The model's recovered regimes matched Currier's 50-year-old manual A/B labels with ARI = 0.383. Of 185 folios, 113 (61%) were assigned to the correct class with over 90% posterior confidence.
Note the striking implication: a cryptographer's naked-eye observation from 1976 was independently rediscovered by a 2026 probabilistic generative model. The data itself is speaking.
But that wasn't enough. Parisel did a fourth thing: predictive validation. He trained a supervised classifier on some folios, then predicted the A/B identity of completely unseen folios. Cross-validation accuracy: 89.2%. Of 500 random label permutations, zero achieved comparable accuracy (p < 0.002).
This is not description of training data—it is prediction of future data. The A/B distinction is not a retroactive labeling game; it is a real, predictable statistical property.
5. Three Functional Zones: The Cipher's Fingerprints
Parisel's deeper analysis revealed a finer structure: character pairs split into three functional zones by their "loyalty" to the A/B distinction.
Categorical Zone
Includes d/l, or/ar, s/r, e/ee. Their Cramér's V exceeds 0.20 (a strong association), they show large regime differences in the mixture model, and they are the pillars of the clustering.
These pairs behave exactly like cipher substitution: in A, almost only one variant is used; in B, almost only the other. The switch is near-binary—like a switch: on or off, no middle state.
Intermediate Zone
Includes ol/al, y/dy, k/t. Cramér's V between 0.04 and 0.15, moderate PC1 loadings. Not absolute like a switch, but they do carry A/B boundary signal—more like graded preferences: one variant favored in A, the other in B, with occasional exceptions.
Free Variation Zone
Includes o/a, ch/sh, f/p. Cramér's V near zero, PC1 loadings near zero, minimal regime differences in the mixture model. Completely unconstrained by the A/B distinction.
The existence of these three groups is profoundly significant. They tell us the Voynich Manuscript's "writing system" is not a simple, globally uniform cipher table. Some pairs are strictly regime-dependent (like cipher substitution), some merely preferential, some entirely unaffected (like genuine allographic variation).
6. The e/ch Anomaly: A Rebellious Pair
One character pair fascinated Parisel especially: e/ch.
In aggregate statistics, e/ch has Cramér's V of nearly zero (0.007)—apparently a "free variation" member. Yet at A/B boundaries it shows the largest jump (p = 2.45 × 10⁻⁶). Strangely, including e/ch in clustering features actively suppresses clustering—unsupervised K-means ARI drops to 0.208; removing e/ch doubles ARI to 0.456.
Parisel's explanation is elegant and deep: e/ch variation is not governed by the A/B binary regime, but conditioned by more local factors—possibly word-level structure or character position within words. It carries the most local boundary information, yet is most disruptive to the global partition.
Like a witness in a detective story who knows the most about certain local details, but whose full testimony would collapse the entire case's framework. e/ch reminds us: the A/B distinction, while important, is only the tip of the manuscript's statistical iceberg.
7. The Explained 29% and the Unexplained 71%
Parisel's model yields a sobering quantitative result: the A/B labels explain 29.3% of the variance in character-pair ratios across folios (R² = 0.293).
What does this mean?
Currier's A/B distinction is indeed the manuscript's dominant statistical axis. But it explains less than a third of the variation. The remaining 71%—more than seven-tenths of the statistical structure—comes from sources the binary A/B framework cannot capture.
Parisel offers clues:
8. A Blow Against the "Hoax Theory"
Voynich studies has a persistent debate: is it all a hoax?
Gordon Rugg proposed in 2004 that the entire manuscript could be meaningless text generated by a 15th-century charlatan using a Cardan grille. Andreas Schinner and Timm further proposed possible generation algorithms in 2020. If the manuscript is merely well-designed gibberish, it doesn't need "deciphering"—it has no meaning.
Parisel's results strike a heavy blow at hoax theory.
If the whole book were generated by one person, one method, in one continuous creative process—hoax or cipher—its statistical structure should be uniform, with at most random fluctuation. But Parisel demonstrated:
1. A single uniform source cannot generate the observed A/B differences (200 simulations, all failures). 2. The two regimes differ not only globally but are detectable at internal boundaries within the same quire (p = 6.89 × 10⁻⁵). 3. The character-pair switching patterns show a cipher-like hierarchical structure (categorical/intermediate/free variation)—a feature random text would not naturally produce.
Hoax proponents can argue: "the author deliberately designed two gibberish systems to enhance mystery." But the cost of this ad hoc hypothesis is enormous—you must assume a 15th-century author was not only versed in statistical patterns but deliberately created a structural complexity that even 21st-century machine learning models need five pages of mathematics to approximate.
The more parsimonious explanation: the manuscript contains real, meaningful structure. It may be a cipher, an unknown language, a constructed language—but it is unlikely to be pure gibberish.
9. A Dialogue with the Naibbe Cipher
Just months earlier, science journalist Michael Greshko published a fascinating "counter-proof" in Cryptologia: rather than deciphering the Voynich Manuscript, he reversed the problem, designing a cipher feasible with 15th-century technology—the "Naibbe cipher"—that encrypts Latin or Italian into ciphertext statistically very similar to Voynichese.
The Naibbe cipher's core is verbose homophonic substitution: using a deck of cards or dice as a randomness source, mapping plaintext letters to multiple different Voynich-like symbol strings. Greshko showed this method simultaneously reproduces several key statistical features: symbol frequency, word-length distribution, positional patterns.
But the Naibbe cipher cannot perfectly reproduce Voynich B's properties. This means: even if the cipher hypothesis is correct, the real system is more complex than Naibbe.
Parisel's A/B research provides new constraints for cipher theory. Any candidate cipher must explain:
10. Closing Thoughts
Parisel's paper title is modest: "A Quantitative Confirmation of the Currier Language Distinction." But it does far more than "confirm."
Using five independent lines of evidence—label permutation, Markov simulation, quire boundary analysis, unsupervised mixture modeling, predictive validation—it builds an almost unassailable argument: the Voynich Manuscript genuinely contains two statistically separable "writing regimes."
It reveals three functional zones, setting constraints on all future decipherment attempts.
It quantitatively tells us A/B is 29% of the story; 71% remains in the dark.
A metaphor: the Voynich Manuscript is not a castle without an entrance. It is a castle whose doors we know exist but whose keyholes we haven't found. Currier pointed in a direction fifty years ago, saying "this side and that side look different." Parisel proved with statistics that Currier was right—and quantified the degree and structure of that difference.
But what lies behind the door? A vault, or a blank wall? We still don't know.
Perhaps we never will. But it is precisely this state of "knowing some, not knowing more" that keeps the Voynich Manuscript haunting us six hundred years later. It is not an unsolved mystery—it is the structure of mystery itself.
---
*This article is based on Christophe Parisel's paper arXiv:2604.25979. Historical background draws on Mary D'Imperio's *An Elegant Enigma* (NSA, 1978), Kennedy & Churchill's *The Voynich Manuscript* (2004), and Bowern & Lindemann's *The Linguistics of the Voynich Manuscript* (Annual Review of Linguistics, 2021). The Naibbe cipher discussion references Michael Greshko's research published in Cryptologia (2025).*