$ styleprofile build examples/writer --contrast examples/llm-drafts -o '<tmp>/writer.json'
[exit 0]
--- stdout
STYLE PROFILE   7 chunks, 4,496 words

As a reference   held-out Delta 0.82 typically, 95% under 0.97 (7 documents)
  Shorter texts: calibrated at 75 words (124 pieces), 150 words (56 pieces); not at 300 words (28 pieces from 7
  documents, worth 18 independent ones), as a length needs 20 independent pieces from 3 or more documents. Under 75
  words, no verdict.

Contrast: LLM drafts   5 chunks from 5 documents; separates them from the reference with AUC 1.00 (perfect separation on
7 reference and 5 contrast documents)
  LLM-likeness on held-out text: reference 0.20 typically (95% under 1.03), LLM drafts 13.72
  Length alone: AUC 0.73 (reference 640 words per chunk, LLM drafts 624): some length difference
  Compared with the reference, LLM drafts have: Long words (7+ letters) ▲, List items ▲, Bold phrases ▲, Headings ▲

Pass --all, or run `styleprofile show` on it, to see the metrics.

wrote <tmp>/writer.json
Next, score a draft against it:
  styleprofile score <draft> <tmp>/writer.json
--- stderr
Thin reference: 7 chunks, 4,496 words (aim for 15+ chunks and 20,000+ words); add more of the writer's documents
