Compliance checks that happen manually before a release are regularly postponed. Better: make them an automated pipeline step.

What can be automated

  • Model registry checks: Is a purpose documented for every model?
  • Data provenance: Are training data sources documented?
  • Documentation gates: Does the technical documentation exist in the expected structure?
  • Transparency: Are user notices for limited risk present?

Example: a check as a script

# Checks whether metadata exists for all models
compliance-cli check --models ./models --require purpose,owner,risk-class

The exit code decides whether the pipeline succeeds. This turns compliance into a measurable quality attribute.

Cultural effect

Automated checks shift the discussion from “Who is to blame?” to “Which rule is missing?” That is the real gain.