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

[nablator]

The autocorrelation is (positive-negative)/(positive+negative).

I get 9.54% on Takeshi Takahashi transliteration lines (global)

9.15% on Currier A paragraph lines of RF1b-er all merged in a single line
9.99% on Currier B paragraph lines of RF1b-er all merged in a single line

By the way this property is related to the clustering of t/k gallows, causing much longer sequences of the same gallows than in a word-shuffled text, itself related to other local similarity metrics.

Any algorithm that scales the probability of picking a source for copy with distance should get these statistical results more or less right.