The Voynich’s Two Voices – A Field Guide to A and B

Two pages apart, the script behaves like two different systems. In the Herbal section of the Voynich Manuscript, you can turn a leaf and watch the statistics flip; word-final ‘dy’ suddenly floods the lines, while the ‘chol’ and ‘chor’ combinations that were everywhere a page earlier almost vanish.

The handwriting changes with it. That is the problem this piece follows.

The Cryptographer’s Ghost

In 1976, at a conference on medieval manuscripts, a US Navy cryptographer named Prescott Currier announced a finding that fundamentally altered the study of the Voynich Manuscript. For decades, the codex had been treated as one continuous text in a single invented script. In his declassified NSA presentation, Currier stood up and said there were two.

His claim was not based on the subjective feel of the writing. It rested on statistics. When he counted letter groups across the herbal pages, the patterns split in half. On pages he labelled as A, certain clusters such as ‘chol’ and ‘chor’ turn up with high frequency. On B pages, the end of words is where the action sits. The bigram ‘dy’ at the word-final position appears over and over. The gap was too sharp to dismiss as chance.

Currier gave them names: Language A and Language B.

The terminology stuck. His point was clear. The codex is not one stream of gibberish dressed up as words. It is two statistically distinct systems occupying the same vellum. This discovery carried weight because of who he was. Currier had spent his career breaking codes during the Cold War, trained to spot the fingerprints of language in hostile ciphers.

Two words act as clear flags. On A pages, the most frequent token is ‘daiin’. On B pages, the champion is ‘chedy’. Currier and later analysts pointed out something sharper still. The word ‘chedy’ does not appear in the A corpus at all, while ‘daiin’ is present in both, an observation detailed in later computational analyses. That asymmetric relationship is a clue we will return to.

For Currier, the implication was obvious. A crude hoax could have been built from random squiggles. But this level of structured difference inside a single book pushed the odds against that explanation. In his words, the dialect split made the possibility of a one-man hoax ‘much less likely’.

A Map of the Fracture

The physical book matters because the split does not track cleanly with its subject matter. The manuscript is a codex of 102 parchment folios, radiocarbon dated to the early fifteenth century, with evidence of missing leaves and later binding work.

Drawings divide the book into broad themes: Herbal, Astronomical, Balneological (depicting bathing scenes), Cosmological, Pharmaceutical, and a dense final section of recipe-like paragraphs. You might expect a linguistic mode to stick to a theme. It often does not.

The Herbal section is the main example. It holds long stretches of Voynich A, then patches of Voynich B, sometimes on facing pages. The Balneological section is different. It is consistently written in Voynich B throughout. The Astronomical pages lean towards A, while the Recipes lean to B. The Pharmaceutical pages are mixed, a distribution mapped in detail on sites like voynich.nu.

The stranger feature appears when you stop reading in the current page order and start tracing the book by its physical units.

In the fifteenth century, a book was made of quires. Folded vellum sheets sewn together into gatherings. A scribe would typically copy one quire after another until the book was finished. That rhythm makes Currier’s finding stand out even more.

The split between Language A and Language B does not follow these themes neatly. In Quire 4, for example, a bifolium written in Voynich A by one scribe is wrapped around another written in Voynich B by a second scribe. This method of assembly is, according to manuscript expert Dr Lisa Fagin Davis, ‘utterly atypical’ for the period. Parallel copying in that period tended to allocate whole quires to different scribes, which kept text blocks intact even if people worked at the same time.

There is physical support for later disturbance.

A large water stain runs across a path that makes little sense in the present folio order and appears to predate the modern folio numbers. If the quires were disbound to dry and then reassembled without original page marks, bifolia could easily have been shuffled in a way that hid the original sequence. That does not explain the dialect split by itself, but it does explain why the split looks chaotic in places.

The Hands Behind the Split

The statistical divide Currier found is matched by a physical one – the handwriting.

Currier himself first noted two distinct styles. On the pages he labelled Language A, the script was neat and widely spaced. On the pages he called Language B, it was tighter, more slanted, and less consistent. To his eye, two different people were writing. He called them Scribe 1 and Scribe 2.

More recent and detailed palaeographic analysis by Dr Lisa Fagin Davis has confirmed and expanded this picture. Applying the principles of studying historical handwriting to the script’s letterforms and habits, she identified five distinct scribes across the codex.

The key point is where those five land on the A/B divide. Scribes 1 and 4 write in Voynich A. Scribes 2, 3, and 5 write in Voynich B. No single hand has been shown to write both dialects. The visual differences are not minor. Scribe 1 has neat, even lines that sit flat on the ruling. Scribe 2 presses text upwards in a cramped slope. The clustering of hands by dialect is the strongest link we have between the statistics and the physical object.

This was not a random division of labour. In medieval scriptoria, multiple scribes often shared out portions of a book. What we see here is stranger… a group of scribes copying within one project, yet split down the middle into two incompatible systems of writing.

The Contradiction Log

Several tensions in the record drive the rest of the debate. Each one forces a choice about how the book was made.

• One book, two systems: A physically coherent codex, with consistent vellum and pigments, shows two distinct statistical systems in its text. Both are internally consistent, yet they share a large amount of common vocabulary.
• Sharp split, soft edges: The A/B contrast is clear in many places, yet a handful of folios sit between the poles on standard measures. These intermediate pages look more like a transition or a continuum than a hard wall.
• Topic does not rescue it: If dialects mapped to subject, the Herbal section would not mix A and B so intimately. It does. The Balneological section, being uniformly B, shows that content can correlate locally, but it is not a general rule across the book.
• Five hands, two dialects: Several people wrote this book. They cluster into two dialect families, not five individual styles. That implies shared rules and coordination, not just personal habit.
• Vocabulary asymmetry: Voynich B seems to contain almost all of A’s common words plus sets of forms that A never uses. ‘chedy’ is the poster child for a B-only high-frequency word. If A and B were two unrelated languages, we would expect sharper, more symmetrical divergence.

Weighing the Working Theories

By now, the pattern is clear. Two statistical ‘languages’, five scribes, interleaved bifolia, and a few pages that sit between A and B. Three explanations are usually put forward. Each solves part of the puzzle and strains against another part.

Theory A: Multiple Authors

• Claim: Different scribes used different norms of Voynichese, producing the A and B dialects.
• For: The handwriting evidence is hard to argue with. Five distinct hands have been identified, and they cluster by dialect. Scribes 1 and 4 write in A; Scribes 2, 3, and 5 write in B. No hand crosses over. That is the cleanest line in the whole manuscript. It fits what we know of fifteenth-century workshops, where several scribes often split a job. If you start from the page, not from the statistics, this is the model that matches what you can actually see.
• Against: The language data do not split so neatly. Voynich B looks like a superset of A. Words common in A also occur in B, but B carries additional forms that A never uses. That asymmetric relationship is hard to square with two unrelated systems used independently by separate people. It implies a shared rulebook or a sequence in which one system expanded on the other. If five people produced only two tightly bounded styles, where did that shared constraint come from, and who enforced it?
• Verdict on A: Best fit for the visible hands and the practical reality of a workshop, but it needs a coordination mechanism to explain the vocabulary asymmetry and the discipline across multiple scribes.

Theory B: Evolving Method

• Claim: One author, or a single guiding intelligence, changed the way the text was generated partway through. In simple terms, the system evolved from A into B.
• For: The vocabulary asymmetry points this way. If B extends A, an evolution is exactly what you would expect. The ‘continuum’ pages that sit between pure A and pure B also make sense as a transition. B pages are typically more verbose, which could be the by-product of new rules or tokens added to the generation method. If you treat Voynichese as a cipher system rather than an unknown natural language, an incremental change in rules neatly explains those statistics.
• Against: The handwriting does not follow. Scribe 1’s neat, horizontal hand writes A. Scribe 2’s cramped, slanted hand writes B. If this were a single-person changing method, we would still expect the same hand to show up on both sides of the transition. It does not. The bifolium-by-bifolium distribution also suggests parallel work by different people rather than one writer moving forward in time with a new system. In short, the palaeography breaks this theory’s spine.
• Verdict on B: Best fit for the statistics, but it collides with the fact of multiple, stable hands tied to one dialect each.

Theory C: Content-Specific Language

• Claim: The dialects were chosen by topic or function. The scribes switched mode when the subject changed.
• For: Some sections are internally consistent. The Balneological pages are all B. Several recipe pages are B as well. Information theory studies have shown that certain keywords cluster by section, which shows some coupling between words and pictures. If you only sampled those areas, you might believe the topic drives the choice of dialect.
• Against: The Herbal section breaks the model. It is one topic split across both dialects, sometimes on facing pages, sometimes interleaved by bifolium. Parts of the Pharmaceutical section show the same problem. If subject matter drove language choice, Herbal should not be mixed. It is. That makes the topic a weak predictor at best.
• Verdict on C: Explains local pockets of uniformity, but fails on the biggest section of the book.

Sources

Sources include: Prescott Currier’s declassified NSA presentation on the A/B split (1976); the extensive collation data, language pages, and transcription resources curated by Rene Zandbergen; palaeographic work by Dr Lisa Fagin Davis identifying five distinct hands and analysing the manuscript’s construction; the Yale Beinecke Library’s digital collection and catalogue entry for MS 408; and published computational analyses of word co-occurrence and vocabulary asymmetry.