List of "weird" vords

[oshfdk]

I would expect that each page would contain some filler text and some real text with the larger proportion being filler. So, I doubt there would be any pages with just filler text and no real text. Likewise, I doubt there would be any pages with just real text and no filler text.

So, I suspect the gibberish would be interspersed with real words. 

I don't think this changes much as long as there is more gibberish than words. Gibberish generated without some proper randomization would probably drive repetitions up.

The question of how gibberish text was produced is akin to the question that people who argue the manuscript is a hoax and therefore all gibberish have to address.

Agree.

If you can provide statistical evidence for this that would be interesting.

Ok, suppose some of the character sequences in the manuscript are the filler and some are the actual ciphertext. This means that character sequences would be sampled from two different distributions with different properties. Let's try a simple experiment, take 1000 most common substrings (character sequences of any length, including dots, commas, etc) from ZL EVA transcription, and plot a bar chart of their repeat counts (logarithmic to make the data easy to see).

Then repeat the same experiment with Opus Majus and then take Opus Majus and replace 60% of words with random gibberish of the same lengths.

In the image below you can see three resulting charts (bar colors code the lengths of corresponding substrings, and each 100th substring is shown below the chart just for a bit better exposition of the data). And while the Voynich and Opus Majus charts look quite similar in shape (not in color, though, highlighting unusually strict structure of Voynichese with many repeating sequences of 5-6 characters), the chart that combines Opus Majus with truly random gibberish looks quite different from both.

   

This doesn't really prove anything, but I expect that in most cases if the Voynich MS was created as a combination of two distributions - meaningful and meaningless - this would have been quite within reach of modern computational analysis for many years now, and someone somewhere would create a reasonable split of Voynichese into two statistically dissimilar distributions according to some simple rule (as you said yourself, there should be a rule governing which parts to read and which to skip).