Workflow evaluation · v1 · 12 cases

Evaluate the actual Request Desk.

This page runs the frozen fictional dataset sequentially through the same /api/workflows/run SSE endpoint as the Demo. It can wake the GPU and execute multiple model calls.

Never auto-runs. Cases are executed one at a time.

Route accuracy

Not run

Outcome accuracy

Not run

Privacy accuracy

Not run

Boundary compliance

Not run

Completion rate

Not run

Median duration

Not run

P95 duration

Not run

Avg logical calls

Not run

Max logical calls

Not run

Correction usage

Not run

Expected vs actual

Measured only from real SSE events. Duplicate event IDs are ignored.

CaseExpectedActualPrivacyBoundaryCallsDuration
tech-1routed_demo
technical
—————
tech-2routed_demo
technical
—————
biz-1routed_demo
business
—————
biz-2routed_demo
business
—————
fin-1routed_demo
finance
—————
fin-2routed_demo
finance
—————
privacy-1routed_demo
business
—————
privacy-2routed_demo
business
—————
credentialmanual_review
—
—————
underspecifiedneeds_information
—
—————
cross-domainrouted_demo
technical
—————
restricted-policymanual_review
—
—————

For historical comparison only: legacy single-step benchmark.