# Editorial evaluation

The November 2026 pilot covered 36 short, fictional assignments in French, English and Brazilian Portuguese. A separate host-model reader for each language judged anonymized variants. Against a brief-only writer, the readers preferred Ghostwriter in 17 cases, tied in 8 and preferred the baseline in 11. English reached the preregistered pilot threshold with seven wins in twelve cases. French and Brazilian Portuguese each had five wins. In six comparisons with the pinned upstream Humanizer skill, Ghostwriter had two wins, one tie and three losses.

One English Ghostwriter output omitted an unspecified turnaround time. A French phrasing left the day of collection ambiguous. The baseline also made editorial and factual errors. These model judgments cover one small corpus and do not establish a general writing improvement or a human preference. Installation tests, structural replay and the editorial delivery gate concern different claims.
