TEVV Throughout the AI Lifecycle

Test, Evaluation, Verification, and Validation (TEVV) are distinct but complementary activities that should occur throughout the AI lifecycle, not just at deployment gates.

Test: Examine system or components to detect problems Evaluation: Assess performance against requirements Verification: Confirm the system was built correctly (meets specifications) Validation: Confirm the right system was built (meets user needs)

Key structural insight: AI actors performing verification and validation should ideally be distinct from those performing test and evaluation. Independence improves rigor.

Phase-specific TEVV tasks:

• Design: Validate assumptions, data collection decisions, measurement approaches
• Development: Model validation, algorithm assessment, bias evaluation
• Deployment: System integration testing, compliance verification, user experience evaluation
• Operations: Ongoing monitoring, incident tracking, emergent property detection, model recalibration

TEVV as regular process, not checkpoint, enables mid-course correction and post-hoc risk management.

Related: 05-molecule—govern-map-measure-manage-framework, 05-atom—ai-risk-measurement-challenges, 05-molecule—distributed-responsibility-ai-actors