Wordsmith Model Bench
What it measures and how
October 2026 pilot: 12 model configurations across craft notes, line edits and polish, scored 0–100 with 95% intervals. Judges from different model makers never grade their own maker's models.
How this evidence is used
Under observation, not scored: there is no data reuse license and no machine-readable results, and Wordsmith, which sells a manuscript editing product, writes and grades the test itself. Instructions, checklist and pages are unpublished, so results cannot be checked. Reassess once a license is stated, machine-readable results are published and the judges are validated against human editors.
Limitations and attribution
Described by its operator as a pilot, not an independent certification. Small sample with wide intervals; judges not yet checked against human editors; English manuscripts only.
Data license: No data license is stated on the page.
Results are published by Wordsmith. Raw scores and the e/acc consensus index use different scales and cannot be added directly.