The structure of the Voynich text and how it may be generated

[magnesium]

I’m so interested in seeing how this progresses; congratulations. Adding a mild spacing ambiguity to your model as a final gloss is probably enough to boost the number of hapax legomena. For example, at the very end of generating a 38,000-token text, you could randomly omit 2-3% of all spaces, representing either scribal cramming or modern mis-transliteration. You could even add a line-position bias in that, where the probability of dropping a space increases from left to right, consistent with the observed left-to-right space narrowing in the VMS.