Two results, each scoped to the corpus it was measured on: 97.9% writing-profile attribution accuracy on a nine-author, 108-document legal corpus, and F1 = 0.935 contributor classification across 66 contributors and 53,000+ comments in the Drupal core issue queue.
What was measured
| Task | Corpus | Scale | Metric | Result |
|---|---|---|---|---|
| Writing-profile attribution | Legal documents | 9 authors, 108 documents | Accuracy | 97.9% |
| Contributor classification | Drupal core issue queue (drupal-core-bot) | 66 contributors, 53,000+ comments | F1 (precision and recall combined) | 0.935 |
How the engine works
The Profile Engine builds an author profile from a corpus — 129 stylometric features per profile — then verifies attribution and monitors voice-fidelity drift. F1 combines precision and recall: it rewards finding the right patterns without flooding reviewers with false positives.
What these numbers do not claim
They do not claim the engine reaches 97.9% on your corpus, on shorter texts, or on more authors. Attribution difficulty scales with corpus size, author count, and text length, which is exactly why each figure here travels with its scope. We report what we measured, not what we hope generalizes.
Checking the work
A fuller protocol write-up is planned; this page is the first installment, published so the figures circulating about the Profile Engine stay attached to their scope. The platform core that hosts the engine's integration is public at github.com/zivtech/joyus-ai. Questions about the methodology: info@zivtech.com.