AI operating governance · Field note

An AI evaluation is a production operating control

Treat evaluation as a living production control for changing workflow behavior, not a one-time technical test.

TL;DR

Treat evaluation as a living production control for changing workflow behavior, not a one-time technical test.

What the paper develops

A one-time evaluation cannot protect a workflow whose users, data, models, and consequences continue to change. This paper treats evaluation as a living production control: a maintained set of representative cases, thresholds, exception review, drift signals, incident learning, and an owner who can change permitted use. The emphasis is on evidence that remains relevant after launch.

The operating move

Run evaluation as a recurring production control. Test the real workflow, monitor drift and exceptions, and change the evaluation set as users, data, models, and consequences change.

WORKFLOWCONTROL EVIDENCEHUMAN OWNER

Inside the white paper

  • A production evaluation set grounded in representative workflow cases
  • Thresholds, drift, exceptions, incidents, and change triggers
  • Ownership and cadence for updating reliance as evidence changes

Sources and notes

  1. NIST
  2. Blake Bullwinkel et al.