Writing a fake Voynich by rules: copies, habits and the link between words

[quimqu]

Yeah, so it’s exactly what I was afraid of—a replica tailored to the text. You can build that with AI in just a few hours—and depending on how closely you copy the original, it can be very accurate. At least if you understand the structure of the VMS.

Of course, at the first level, that’s kind of pointless. Theoretically, you could do the same thing with any normal language.

But given the specifications you describe, something else would be extremely interesting:

What if you “applied” the generator to other European languages with similar basic requirements? Then you’d be able to spot differences in comparison, which in turn would—or could—reveal quite a bit about the VMS.

I think that would be the more exciting part of this.

Or what do you think?

On "tailored": the letter tables are learned from the text, yes. But learned only from the odd folios, it writes the even ones and matches 51 of 64 measures, while the two real halves match each other on 55. So it's not copying the text, it's following the rules.

And the generator is only a tool for me, not the goal. I use it to check whether the rules I measured are complete: if the fake text fails somewhere, a rule is missing or wrong.

Comparing with natural languages would be interesting, but it's harder than it looks. Most of the structure in the VMS lives in the line and the page: line starts and ends, copies from a few lines above, habits per sheet. For a fair comparison I'd need medieval texts transcribed line by line, with the original line breaks and pages, and most digital texts have lost exactly that. I did use a Latin text (Cusanus) as a control for the word-level measures, and there the VMS is clearly different. If someone has a good line-by-line transcription of a 15th century manuscript, I'd be happy to run the measures on it.