A One-Page Ledger Method for Generating Voynich-Like Text

[Dunsel]

I think you did an interesting work and you deserve some compliments, but I fear there are some problems.

You followed Torsten Timm's copy-and-modify algorithm, but with a twist. Torsten enforces the generation of Voynich-like words by having a (complex) set of modification rules. Instead your modification algorithm is simple (I did not check it, but I'm sure it is) and you enforce the Voynich-likeness of the words by adding the ledger sieve at the end. So, the basic idea is Timm's, and you are proposing an alternative approach in its implementation. Which is surely a good idea: a drawback of Timm's implementation is that his modification rules are many and complex, and with enough rules (enough parameters) one can approximate any text to any arbitrary degree of precision. And in my opinion, Timm's implementation uses too many rules to achieve too rough an approximation of the VMS word structure, so it's not a smoking gun for the copy-and-modify theory.

Your approach is surely interesting, but ultimately it runs into the same problem. You can use very complex 'ledger tables' (many parameters) or, as intelligently you choose to do, you can use simpler tables (less parameters), but by doing so you generate distinctly un-Voynich like words, ie. in your sample there are three 'ddocta' in the six rows of text starting at <f3v.P1.2;K>. The sieve has too large a mesh (this is the 'fankao' problem I noticed in a previous post), and to narrow the mesh you would need to add more parameters (many, many more).

But in any case, a nice work. I hope you will go on searching for clues to unravel the VMS!

Thank you.  It's much appreciated.

I actually agree with a good part of your criticism, especially regarding parameter explosion. One of the things I was specifically trying to avoid was the “rule accretion” problem where enough rules can eventually approximate almost anything. In my generator, the mutation engine itself is intentionally simple; most of the constraint comes from the ledger and the source pool rather than a large library of rewrite operations.  My goal has been to model those thoughts in the scribe's head.  "I like this word. I'll use daiin a lot."  "I need a really cool word to start the page so I'll put a gallows there."  The problem with python is, it doesn't understand, "don't make a stupid word," without some sort of weighting. The human eye does that easily. And that's where I had little choice. I've had the analysis for months that told me words came from sheets. But without something to generate text it wouldn't likely be believed or taken seriously.  And, trying to model human behavior in order to create a generator has been a nightmare.  I have at least 100 versions lying around and a dozen, "dammit, that goes in the trash. Start from scratch," days.

Where I think the approach differs from Timm’s is less in “copy-modify exists” and more in the proposed production ecology. The generator is not operating from a global vocabulary but from localized sheet/page source pools, with strong positional behavior, limited working memory, and adjacency legality acting as a scribal aid rather than a linguistic grammar. The analyzer side is actually more important to me than the generated output itself, because Scribe 1 repeatedly collapses into small sheet-level source packets.  To my knowledge, no one has been able to point to a Voynich page and say, all those words can be explained by this sheet.  And maybe one or two from this sheet. And that is what I think I've demonstrated in the paper.  

I think your broader criticism about distribution is fair. I would argue, however, that forms like “ddocta” are not actually un-Voynich; they are valid manuscript forms. The real question is not whether such words can exist, but whether the generator reproduces their behavior correctly. There is no reason a rarer form could not temporarily or permanently become a family center word in the same way that more dominant forms like “daiin” appear to do. The manuscript itself shows evidence of lexical blooms that appear and vanish, and shifting local families. I think that is one of the reasons sheet-level sourcing matters more than fixed global vocabularies. I have no proof but I suspect they would be emergent based on which sheets are used as sources and where.  I had a real hard time with this myself at first. We've seen so much Voynich over the years that we have certain expectations.  Every time I spat out 100 pages of generated text, I went hunting for daiin on that last page.  I had to keep telling myself, look for a family, don't assume it's daiin.  I kept looking at pages thinking, "does that look Voynich?" and I had to instead keep thinking, "look at the numbers".  Hell, in early generator models I was synthetically injecting daiin just so the page would 'look' right.

So I do not see this as “case closed” for copy-modify at all. At best, I think the work shows that a relatively small local system can reproduce a surprising amount of the manuscript’s ecology without requiring a hidden language or a gigantic encoding apparatus. The hard part now is reducing the remaining synthetic artifacts without turning the system into an overfit monster.

Again, thank you!