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

[Mauro]

[I understood each section of the manuscript with its own set of parameters, without mixing languages?

If I understand correctly, the only possibilities are:
→ scribe trained on the whole manuscript
→ scribe that has not seen quire M (run calibrate_section_e.py first)

Yes, it would be better to restrict the training to a set of pages defined for example by npcompl33t's categories, then generate a pseudo-section (or several).

It would be also interesting to see what @npcompl33t's serialization algorithm achieves on @quimqu's text: horseshoe or no horseshoe?