Huh. I probably should have read this earlier.

I have a toy tool for facilitating style imitation and combating adversarial authorship analysis: https://github.com/bhickey/doppelsprecher

Given two corpora, it generates two smoothed ngram models and produces estimates of per-word entropy ratios. If an author can see the highly entropy portions of a text, they'll have the opportunity to change it.

Pardon the license, it'll change to something sane shortly.