The Challenge of Analyzing a Dynamic Text

[Torsten]

One point on this:

The observations on the development of vocabulary in the VMS are real and valuable, but I consider the conclusions drawn from them to be questionable. Having recently studied medieval German manuscripts intensively – Ortloff von Baierland, the Breslau Pharmacopoeia, Admonter Bartholomäus – I have also noticed such deviations within individual texts by individual authors. Spelling changes over the course of a manuscript. The frequency of words shifts dramatically depending on what is being described. The similarity between early and late sections decreases significantly, even in a simple medical text. All this does not require complex systems or even a self-citation process – it is simply what happens when a person writes a long text over weeks or months, possibly copying from different sources, changing topics or gradually changing their writing habits. The assumption that a “static system” should provide identical statistics on every page sets a bar that no real medieval manuscript could reach.

Thank you for this thoughtful critique grounded in concrete manuscript experience. You are right that spelling variation, vocabulary shifts, and evolving scribal habits are normal features of extended medieval text production. The paper does not dispute this.

However, the Voynich Manuscript does not merely show variation — it shows a specific combination of structural properties that medieval manuscripts do not exhibit. In the Voynich text, 84.67% of all word types form a single connected network through single-glyph edits. High-frequency words systematically have more graphically similar neighbors, while words appearing only once are isolated in the network. On more than half of all pages, two of the three most frequent words differ by only a single glyph. Vocabulary evolves in strict directional chains with intermediate forms (<chol> -> <cheol> -> <cheo> -> <chey> -> <chedy>), and words from early sections persist in late sections but never the reverse.

None of these properties characterize natural language texts, however much their spelling varies. In Ortloff von Baierland the word "und" may be extremely frequent, but it does not generate dozens of single-edit variants clustering on the same pages. Function words are distributed throughout the text regardless of topic — unlike Voynich, where the most frequent tokens can change from page to page. The paper does not claim that a real manuscript should produce identical statistics on every page; it argues that this particular constellation of properties observed in the Voynich text requires a different explanation than normal scribal variation.