[Movie] This Famous Medieval Book May Be a Hoax

[Mauro]

Only if the method is deterministic, otherwise you have a one-way cipher.  ...

From what I understand, the self-referential method relies on selecting a previous word and choosing a modification to apply to that word.

The method isn't very clear about how these two choices are made,

The exact method used is in the source files of Torsten's program on GitHub (You are not allowed to view links. Register or Login to view.) if you can read code. I just skimmed through them, what I can say is the choice of the type of modification is a random extraction with parametrized probabilities (comments after '//' are mine):

int combineSplitProbability = 10;   // 10% probability to choose 'combine (rule #3)' or 'split' (more parameters should determine the ratio between combine and split)
int addRemoveProbability  = 30;   // 30% probability to choose to 'add' or 'remove' a glyph or a glyph string, which are defined with more parameters
int replaceProbability = 60;          // 60% probability of replacing a glyph or glyph string (with additional parameters as above)

I did not analyze the code enough to be sure of how the choice of the word to copy is made, but I guess it's chosen randomly (I guess with an uniform distribution) from a sliding window of the text (ie.: from the list of the N preceding words).

I'm also not sure of how many total parameters there are, I'd say at sight 100-300? I think it's a rather low number (which is a good thing), but nonetheless the textual output compares rather well to the original Voynich (which is also a good thing, ofc). To give you an idea of the scale, I wrote a program which can generate a much better pseudoVoynich, but using a whopping ~8000 parameters, which are definetely too many.