Voynich Manuscript Production – What the Physical Evidence Shows

The Voynich Manuscript contains two stable statistical text profiles, while its surviving physical order has been disturbed. The evidence still does not reveal how those profiles were originally produced and assembled.

Fifteenth Century Parchment Analysis

Four samples of the Voynich Manuscript’s parchment were radiocarbon dated at the University of Arizona in 2009. Taken from folios 8, 26, 47 and 68, the samples produced mutually consistent measurements. When combined, they gave a calibrated range of 1404 to 1438 at 95 per cent confidence.

Those folios are not confined to one part of the manuscript. Folios 8 and 47 contain Herbal A material, folio 26 is Herbal B and folio 68 belongs to the cosmological section. The tested parchment therefore crosses both of the manuscript’s principal statistical text profiles.

The four results can also be read separately. The 1404 to 1438 range comes from combining their probability distributions on the basis that the parchment was roughly contemporaneous. Critics of that treatment have pointed out that the individual measurements allow a wider span. That does not show that the skins came from different periods: each measurement is still consistent with an early fifteenth-century date.

Protein analysis carried out by the University of York in 2014 independently identified the parchment as calfskin. The tests do not identify when ink was put on the parchment, who wrote on it or whether the manuscript was produced in one continuous operation.

Voynich Manuscript Statistical Text Profiles

On 30 November 1976, Prescott Currier presented what became the A/B distinction at a seminar convened by Mary D’Imperio. He had found statistical differences which divided the manuscript text into two groups.

D’Imperio later reproduced a clear A/B separation through cluster analysis. René Zandbergen found the same distinction at the level of individual character combinations. The EVA bigram ‘ed’, for example, is almost absent from A pages and common on B pages.

Christophe Parisel tested the division again in 2026, but this time without giving his model Currier’s classifications. Using raw character counts, the Beta-Binomial mixture model selected two groups. It then classified withheld folios as A or B with 89.2 per cent accuracy.

Subject matter does not account for the split on its own. Herbal pages include both A and B. In the biological and stars sections, however, the text profile corresponds more closely with the illustration category. Transcription remains one limitation on all such computational work. Statistical studies depend on datasets in which researchers have had to decide how individual signs should be represented and, in particular, where word spaces occur.

Palaeography and Scribe Assignments

Currier did not regard A and B as statistical categories alone. He associated them with differences in handwriting and proposed that language and hand correlated.

Currier’s A/B classification follows complete bifolia, the folded sheets from which gatherings are formed. It does not change arbitrarily between the leaves or sides of a bifolium.

Lisa Fagin Davis approached the handwriting separately. Her 2020 DigiPal study compared letter forms across the manuscript and identified five scribal hands. Those assignments also run by bifolium.

The claimed connection between handwriting and Currier language is less settled than the existence of the A/B split itself. Zandbergen describes the correlation as visually clear in the herbal section, where Herbal A and Herbal B alternate, but says it has not been sufficiently demonstrated throughout the remainder of the manuscript. That does not show that the correlation fails elsewhere, only that Currier’s broader claim remains unconfirmed.

Davis’s five-scribe classification has also been challenged. In 2026 Torsten Timm argued that some of the glyph features used to distinguish the proposed scribes occur much more widely across the manuscript than the five categories suggest. He reads the handwriting changes as variation along a continuum and proposes that they could represent one evolving hand rather than five distinct writers.

Davis’s study was peer-reviewed, while Timm’s 2026 challenge has not been confirmed as peer-reviewed. No quantified handwriting test currently resolves the disagreement.

The Atypical Manuscript Assembly

Davis found bifolia attributed to different scribes interleaved within individual quires. Quire 4 supplies a direct example: an A bifolium wraps around a B bifolium. She has described this arrangement as ‘utterly atypical’ and ‘extremely rare if not unprecedented’.

There are other signs that the codex has been physically disturbed. The surviving quire numbers omit 16 and 18. Quire 9 is stranger: the sewing now passes through the wrong fold, and its quire mark sits on the first recto.

A later layer of organisation survives in the manuscript’s folio numbers. These were added in ink, probably in the seventeenth century, after the manuscript had already reached a disturbed configuration. Fourteen subsequently lost leaves can still be detected because their folio numbers are absent from the surviving numerical sequence.

The present codex therefore preserves more than one chronological layer: the bifolia on which the text was written, quire organisation, later foliation and subsequent losses. Those layers do not all record the manuscript in the same physical state.

Water Damage and Rebinding Sequence

A water stain runs across the upper margins of the early quires. The later folio numbers sit on top of the staining, so the spill or staining event preceded that foliation.

Davis has proposed a more detailed reconstruction. On her account, a substantial liquid spill led to the manuscript being disbound so that the bifolia could dry. They were then gathered again in an altered order, after which quire numbers were added.

The stain is observable; the spill, disbinding and altered rebinding sequence is Davis’s interpretation of it.

Davis has attempted to use changes in the stain across individual leaves to reconstruct their earlier nesting. She has reported ‘no strong results yet’. Other reconstruction efforts have examined particular gatherings, including Quires 9 and 13 and possible adjacencies among the balneological pages, but no complete original sequence has been established.

That leaves the present A/B interleaving ambiguous: it cannot tell us whether A and B were already mixed when the manuscript was first assembled or whether later reordering brought them together.

Physical Evidence in A and B Profiles

Laboratory analysis has not supplied a material equivalent of the statistical division.

In 2009 the McCrone Research Institute examined manuscript inks and four pigment samples. The tested inks contained carbon, iron, sulphur, potassium and calcium, with trace copper and, in some samples, zinc. Yale conservation findings subsequently described only minor compositional variation and reported iron gall ink throughout the manuscript.

One McCrone sample came from Currier A folio f47r. In the available comparison, it was not materially distinguishable from the tested Currier B samples.

The pigments examined were also compatible with materials available in the manuscript’s period. Blue contained ground azurite with minor cuprite. Green contained copper compounds. Red-brown was associated with hematite red ochre, while the tested white included eggwhite and calcium carbonate.

Multispectral imaging of ten folios in 2014 produced no evidence of underwriting or a palimpsest on those pages.

These examinations are limited in scale. They have not identified different physical production batches corresponding to A and B, but the existing sample is not large enough to exclude subtler variation across the manuscript as a whole.

Manuscript Production Models and Theories

A bifolio-first model is one possible fit for the physical evidence. On this account, the manuscript was written on loose bifolia before binding, with assembly or later reassembly producing the interleaving now visible. The fact that both the Currier profiles and the proposed scribal assignments follow the bifolium is consistent with that model. There is no surviving record showing that this was how the manuscript was actually produced.

Currier proposed a different explanation. His association of A and B with handwriting differences allowed for separate contributors or separate exemplars behind the two text profiles.

Timm and Andreas Schinner have argued in the opposite direction. Their model treats A and B as products of a single text-generation process that changed gradually rather than as evidence for two languages or two scribes. That approach has to accommodate features which do not appear entirely gradual. Parisel’s 2026 analysis describes some character-pair behaviour as effectively categorical even within a broader pattern of statistical variation.

Computational topic modelling published by Rachel Sterneck, Claire Bowern and colleagues in 2021 found clusters associated with a combination of proposed scribe and subject matter. That relationship does not establish whether scribes determined subject, subject affected vocabulary, or both arose from another feature of the manuscript’s production.

No surviving production document identifies the writers, explains the A/B distinction or describes the procedure by which the bifolia were composed and assembled.

Source Box

Sources include: University of Arizona AMS radiocarbon analysis of four Voynich parchment samples (2009); Mary D’Imperio’s The Voynich Manuscript: An Elegant Enigma (NSA/CSS, 1978) and her cluster-analysis work; René Zandbergen’s Extension to the Currier languages; Lisa Fagin Davis’s How Many Glyphs and How Many Scribes? Digital Paleography and the Voynich Manuscript (Manuscript Studies 5.1, 2020); McCrone Associates’ Materials Analysis of the Voynich Manuscript (Project MA47613, 2009); Christophe Parisel’s A Quantitative Confirmation of the Currier Language Distinction (2026); Torsten Timm’s One Hand, Five Labels (2026); Torsten Timm and Andreas Schinner’s A possible generating algorithm of the Voynich manuscript (Cryptologia 44.1, 2020); Rachel Sterneck, Annie Polish and Claire Bowern’s Topic Modeling in the Voynich Manuscript (2021); Lisa Fagin Davis and René Zandbergen on the University of York protein analysis identifying the parchment as calfskin, with the original laboratory report not directly reviewed; and Medievalists.net, Hidden Secrets of the Voynich Manuscript Revealed Through Cutting-Edge Scans (9 September 2024), reporting Davis’s analysis of the 2014 multispectral imaging, with the original Lazarus Project imaging records not directly reviewed.