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

[npcompl33t]

I would like to hear a comment by @npcompl33t (I'm really not in the position of analyzing MDS graphs...). I think that if the seriation on your generated text indeed finds a horseshoe similar to Voynich's, when it shouldn't because the 'sequence' of your text is essentialy random, then this cast doubts on @npcompl33t's results (or there's something else deeper which causes this effect, I don't know). Iirc you also tested @LisaFaginDavis LSA algorithm on your text, and you found it improves the scores even more than on VMS, where it shouldn't, too. But @npcompl33t's results essentialy agreeing with @LisaFaginDavis You are not allowed to view links. Register or Login to view. of Quires 13 and 20 would then be a massive coincidence, probabilities can use different assumptions which will lower the numbers but, face value, they are 1/720 for the exact match on the six folios of Q20 and 1/30 for the reverse match (and an inversion) on the five of Q13, so a total of 1/21600.

So I'd really like to know what's actually going on here. I hope @npcompl33t is reading us too.

This is easy to explain. LFD hands are basically each discrete points on the gradient, as you can see on this chart
   

LFD hands essentially are a weak approximator of the gradient. So if you trained on them as distinctions it would learn differnt parts of the gradient. Since each scribe is mostly active during a single part, with one exception, in 3a. I strongly expect you could train it on the gradient values and have it reproduce the hands. So your model reproduces the chronology, because you are feeding a blurry approximation of the gradient by separating by hand.