Phase 05 — Measurement and Trust
Capability is not trust. Trust comes from evidence that matches the product, calibrated uncertainty, known failure modes, security boundaries, interpretable investigations, accountable ownership, and a release process that can stop or roll back.
Mental model
Trust is a maintained evidence state, not a model property. Measurement, interpretation, security, and governance answer different questions and must converge on explicit release and rollback decisions.
Roadmap through the branches
Exit criteria
You can design component and end-to-end evals, quantify uncertainty, avoid benchmark overclaiming, distinguish explanation from faithfulness, threat-model model/tool/data boundaries, and map governance claims to controls and evidence.
Connects to: Phase 04 — Context and Agency · Phase 06 — Product and Operations
Core sources
- NIST AI RMF — lifecycle risk functions.
- OWASP Top 10 for LLM Applications — application security risks.
- Model Cards — intended use, evidence, and limitations.