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

[quimqu]

What worked was making the change follow the same two letter rule as new words.

Does it mean that you collected statistics of letter trigrams by "picking the habits of the sheet" ? I did exactly that in my first generator. The probability of trigrams including across word breaks were preserved. The first and last bigram of the line were special, creating a simple LAAFU effect because I used a special character for start of line and end of line in the trigram statistics. Now I think I should add a few guardrails to ensure that not too many malformed words get generated: too long or with several gallows...

Not really. The letter statistics are one single table for the whole text (one for A, one for B): the next letter given the two before, also across the space. That table is the same on every page.

The "habits of the sheet" are something separate and much smaller. Just five choices: k or t, ch or sh, q in front or not, dy or y, and words with e or without. Each sheet gets its own preference for each one, the pages move a little around it, and it drifts from line to line. So the letter table says what is possible, and the sheet habit only tilts a few of those choices.