Speculative fraud hypothesis

[Torsten]

If the "copy" and "modify" parts of the algorithm are turned off, leaving only a "generate from scratch" step, it reduces to a Markov chain of order zero -- that just outputs words from a fixed probability distribution, without  regard for previous outputs.

With the right "generate" algorithm, such a generator can reproduce some statistical properties of the VMS text, such as Zipf law, per-word word entropy, word structure, and glyph and glyph pair distributions,  But it cannot produce any word-pair distribution other than (random first word) x (random second word).  That is, the next-word entropy would be the same as the plain word entropy.

As I wrote earlier, even with a significant copy probability, the word-pair distribution should gradually tend to this state; since each time the source pointer is reset another random word pair would be added to the repertoire of pairs available for subsequent copying.

You are assuming the existence of a fixed seed text from which words are copied and modified. That is not the case. In the self-citation model, the Voynich text functions simultaneously as both the source and the outcome of the copying process. This recursive dynamic—where new words continually draw on previously generated ones—is precisely what the “self” in self-citation refers to. 

This is also why the process differs fundamentally from a zero-order Markov chain. A Markov chain simply samples words from a fixed probability distribution without regard to prior outputs. By contrast, in the self-citation model, the distribution itself is continuously reshaped by the ongoing act of copying and modification. In other words, the mechanism is history-dependent: each step reflects and reinforces earlier choices, creating the asymmetries and local clustering observed in the Voynich text. This way, the gradual evolution from Currier language A to B can be understood as automatic side-effect of the self citation method.

Note: The algorithm requires only a minimal seed (e.g., a single line of text) to initialize. The specific choice of seed is irrelevant; what matters are the procedures for selecting source words from the previously generated text and the rules governing their modification. In our implementation, we used line f103v.P.9 of the VMS as seed—<pchal shal shorchdy okeor okain shedy pchedy qotchedy qotar ol lkar>—to generate a corpus of more than 10,000 words. The resulting text contained 7,678 Voynich words (70%) and 3,156 non-Voynich words (30%). These figures alone illustrate that the self-citation method can naturally yield text with Voynich-like properties.