Pace the pipeline
Match the flow of AI output to the real capacity of the people who must check it — throughput honesty as a safety control.
What it changes
Who can pull it
What it looks like institutionally
Every review stage has a finite catching speed. When output volume exceeds it, the share of errors caught falls mechanically — no amount of reviewer diligence changes the arithmetic of a saturated queue. Pacing the pipeline means governing volume as deliberately as quality: rate-limiting how much AI output can reach consequential action per reviewer-hour, or scaling checking capacity before scaling generation.
Institutionally this looks like: production quotas tied to review staffing, queue-depth alarms that trigger slowdowns rather than backlogs, and honest capacity math in every deployment plan ("who checks this, and how long does a real check take?").
The alert-saturation failure in the Illinois Rapid Safety Feedback case is the canonical counterexample: thousands of maximum-risk flags against a fixed caseload is a pipeline nobody paced.
Addresses: Reviewer saturation · Alert fatigue · Rubber-stamping under load. Test a version of this lever in the PAN Lab.