The Bench

The Bench publishes our model and tool evaluations: the same batteries, rubrics, and scoring we use internally, with the failures left in. One issue per evaluation cycle, when a cycle closes. Scores are never softened for publication; the full scoresheets, rubric, and run conditions stay on file, and the receipts match every number in an issue.

The evaluation work itself is coordinated over Oids by the desk team.