Judged features: rhetorical habits a reader would notice

How to read the cards, the status badges and the four groups is explained on the Comparisons overview; every statistical term links to the Statistical glossary.

The reference judge

Eight features that need a reader rather than a script, scored by a model on a seeded random subsample of posts from each population against a fixed rubric. The judge sees only the title and text. Each question is one multiple-choice answer; the shares below are compared with a .

A second judge, and the full set

Scoring 900 posts with a frontier model is affordable; scoring every eligible post is not the best use of it. So the same rubric was given to a second, far cheaper judge, Jev, TypeSafe's typed decision model, which answers each question with a probability distribution over the options instead of generating text. It scored the same 900 posts first. The table shows how often it chose the same option as Claude on each question, with (agreement beyond what the two judges' answer frequencies would produce by chance). Questions at or above % agreement were then scored by Jev on every eligible post, including the detection-off posts.

These repeat questions already tested above on the reference subsample, on a larger set and with a noisier judge, so they are reported uncorrected and outside the correction family. Read them as a check on the subsample, not as new findings. The test is over the three labelled groups; the detection-off row is shown for comparison.