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

[oshfdk]

Let's look at some of the things I've had to consider and would also have to be considered in your model.

• Current line position - Where are we on the line?  Do we need a gallows token here?
  
• Current page position - Do we need to create a new paragraph?  Does this need to be a short line to precede a paragraph on the next line?
  
• Selected source token - Which of those 20 words are we going to choose?  Or, are we going to choose a word on the page we just wrote?
  
• Selected mutation type - Are we going to copy the word or mutate it?  We need a probability.
  
• Copy vs mutate probability - What percentage of words are copied and which are modified?
  
• Glyph insertion/deletion weights - When a word is chosen to modify, what letter can be deleted?  What letter can be inserted or swapped for another.  
  
• Length-preservation bias - When we mutate, do we need to keep the same word length or can we shorten or lengthen it?
  
• Repeat-nearby-family bias - Do we need to make this word look like another word we just wrote down?
  
• Avoid-too-many-new-forms bias - Have we used a word and it's mutation too many times already. Do we need to find a new word? 
  
• Proposed token - Ok, we have chosen the word and we're ready to write it.
  
• Rejection reason - Does the word look stupid? Does it have 4 vowels? 5 consonants? Have we repeated the word 4 times?  Does it look like a possible word with vowels and consonants?  Accept or reject and start over.
  

All of those rules, you could look at the page and know almost instantly how to proceed.  Software needs numbers. Weights, probabilities. So could your proposed method be modelled?  Yes.  Would this be easy?  Oh, I can say for certain it would not.

Ok, we can add the line/page relative/absolute position, this sounds reasonable. 

As for the rest of the list, mostly it's ok, except the generic "Does the word look stupid?" and also "Have we used a word and it's mutation too many times already." - absolutely not. I don't think we can expect the scribes to keep the tally of all the words they have written. If some word occurs 2-3 times in the sample of 20 tokens, it can be excluded of course, otherwise it's unreasonable to expect the scribe to keep the count of all the tokens used.

Other than that, the rest of the things you mentioned can be easily implemented in the function, as the restriction allowed for reasonable size of immutable data, like a table of probabilities. If the choice function includes the past 4 tokens, the repetitions can be checked of course, but these should definitely come from the budget of 20 tokens. We can't expect the scribe to scan the manuscript for any extended time before writing each word.

If the method doesn't realistically model the limitations that a human scribe would face, I don't think it can be used as an argument for or against anything related to the actual MS. There are way too many possible methods to generate something similar to Voynichese.