Why and how the text could be Bavarian

[Jorge_Stolfi]

And exactly the same applies to the sh / k / ch phenomenon. The same dependency does not imply the same step-by-step selection structure. This refers to the linguistic level; when read backward, the long-range effect—although mathematically present—is easier to “explain” because it is cumulative—just as with “oq,” only here it spans multiple glyphs, yet still without a “jump.”When reading backward, the (as mentioned, existing) structure is not noticeable; it blends in with the other distributions of the glyphs. See the diagram:

I thought that I had understood your statistics, but now I am confused again.

That diagram seems to be a first-order Markov chain: a mathematical text generation device that can be described as a graph where each node ("state") is the last letter that was generated, and there is an arrow ("transition") from each state X to another state Y labeled with the probability (or observed frequency) that the letter Y occurs after the letter X.  Is that a correct description of your diagram?

Typically people will handle the word space SP as a letter.  Then the number in a transition from any state X to SP is the probability that the word will end after reaching the state X.  Conversely the number on a transition from SP to any state Y is the probability that a new word will begin with the letter Y.

But it seems that you are ignoring the spaces, so a transition from X to Y may be inside a token or across a word space, between two consecutive tokens.  How do you record which is which?  In the second case, the number on the transition would be the probability that a word that ends with the letter X will be followed by a word that begins with the letter Y.  Is that so?

Such a model can give information about two very different processes, but only for one of them will the information be reliable.  We can discuss  that after I understand what your model is, precisely.

All the best, --stolfi