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

[Torsten]

I prefer to analyze the text the scribe actually wrote rather than speculate about the texts he didn't write — since that number is endless.

I will politely disagree with this statement. A model does not just have to explain what the scribe wrote. It also has to explain why the text did not fall apart into garbage.  A lot of objections in these discussions are really variations of “why don’t we see this?” or “why didn’t the scribe do that?” They are questions about constraint.

In my own experiments, most copy/mutate systems fail pretty quickly unless the constraints are tight enough. They drift, bloat, repeat badly, or start producing obvious nonsense. And, given enough generations, even my generator has issues. That failure is informative. It tells us the Voynich text is not just “anything goes” mutation. So I think the unwritten text matters too. If a proposed method could easily generate endless bad Voynichese, then we need to explain why the manuscript consistently avoided doing exactly that.

If we don't explore the possibilities of what wasn't created then we'll never have acceptance for what is created.

You raise a fair point — but it's a different question from the one I answered. Zandbergen asked why some unspecified modifications are absent. I said: because the scribe took one path and not another, and we can only analyze the path taken. That stands.

Your question is different: why doesn't the text fall apart into garbage? That's a question about constraint, and it has a straightforward answer.

The copying process inherits structure from the source. Each new word is a modification of an existing word that already has valid structure. A modification of a valid word almost always produces another valid word — because the stroke-level constraints are preserved by the modifications taken. The structure doesn't need to be maintained by external rules — it's carried forward by the copying process itself.

And the scribe has eyes. If a modification produces something that looks wrong, he sees it and doesn't write it. Your own "don't look stupid guardrails" describe this exactly. The scribe isn't following a ledger. He's looking at what he produced and judging whether it looks acceptable. Human aesthetic judgment is the constraint. 

His visual judgment follows the stroke-level structure of the writing system. Schwerdtfeger described in 2008 four design rules: (1) line-glyphs can follow line-glyphs or 'a'; (2) curve-glyphs and 'a' can follow curve-glyphs; (3) the 'l'-glyph can be used as a curve-glyph or as a line-glyph; and (4) gallows glyphs count as curve glyphs (see Timm & Schinner 2020, p. 10). These aren't rules the scribe memorized — they're descriptions of what visual similarity produces. A scribe modifying strokes naturally stays within these constraints because crossing them would look wrong to his eye.

This is also why your generator struggles — and why any generator will. For the human mind it is difficult not to repeat itself. The variation comes from imperfect repetition — not from randomness. The consistency comes from habitual patterns — not from explicit rules. A computer needs both dice (for variation) and rules (for consistency) as separate mechanisms. A human produces both from one source — his own cognitive habits. That's why modeling a human scribe in Python is, as you put it, "a nightmare." You can't model the human mind like a computer — because the human produces variation and structure from the same process, while the computer needs separate mechanisms for each.

The manuscript doesn't fall apart because every word was seen and accepted by a human before it became part of the source pool for the next word. The constraints are the pattern recognition abilities and the habits of the scribe himself.