Is there enough scoped, auditable evidence to design a permissioned Trust & Safety pilot?
Fictional demo. All values and reviews are invented; this is not a GEN result or authorized communication.
In 4 fictional paired tasks, median triage time fell by 18 minutes.
4 paired tasks
4 paired tasks; no generalizable interval is claimed.
Median of each task’s before-minus-after minutes; the same task is paired across conditions.
Descriptive for 4 fictional tasks; no cross-agency or causal claim.
Source versions: SYN-STUDY-01 v1 · Review: synthetic-evaluation-reviewer, 2026-10-02
5 of 6 fictional reviewed cases had a complete source, decision, reviewer and timestamp trail.
6 reviewed synthetic cases
The interval describes this small evaluation sample, not field performance.
Complete reviewed cases divided by all distinct reviewed cases; incomplete cases remain in the denominator.
A sample audit trail is not a GEN deployment-quality claim.
Source versions: SYN-EVAL-01 v1 · Review: synthetic-evaluation-reviewer, 2026-10-02
Agree on one buyer task, a representative adjudicated test set, baseline timing, error costs and permission to publish. The ten-example fictional benchmark in the reviewer queue is held because its uncertainty and deployment fit do not support a buyer quality claim.