In development · Updated 2026-11-03
Why this benchmark exists
Assess whether a system selects a justified next action from business relationship records.
Methodology draft
Methodology draft. Tasks will be built from de-identified business records licensed through SourceX with written supplier consent. Supplier identity will never be disclosed. Experts will write task-specific rubrics. A three-run protocol will document inputs, scoring and variation. Dataset composition, judges and release protocol are not finalized.
SourceX authenticated run
In development. Task construction, dataset composition, expert rubrics and the run protocol must be reviewed before an authenticated run. No results are published.
Limitations
No dataset or results have been released. Planned tasks and rubrics require review.