Do the rules agree with people?
I tested the rules against a public dataset of fiction paragraphs, edited by professional writers and ranked by readers. A script computes every number here from the data.
Loading results.
Which version do readers prefer?
Kappa corrects for agreement by luck. Zero means chance and one means perfect. The second line is its 95% interval, which resamples whole paragraphs.
Do findings land on the words writers changed?
What writers changed
Writers cut stock phrases. They kept or added em dashes and long sentences, which my rules ban and fiction uses.
On texts written to test the rules
The Claude judge
The data
The paragraphs, edits and rankings come from LAMP (Chakrabarty, Laban and Wu, CHI 2025), released by Salesforce under the BSD-3-Clause licence. Only the fiction is used. None of its text is copied here, and readers are numbered, never named.