Appendix B: Glossary
Approval loop: A pre action human review process that must approve an AI action before execution.
Research spine: this chapter stays grounded in NIST AI Risk Management Framework and NIST Secure Software Development Framework, then applies that evidence to the operating judgment in the book. Read this alongside the Human In The Loop book, the AI-Native thesis, and the full book library when you want the surrounding argument. Approval loop: A pre-action human review process that must approve an AI action before execution.
Audit loop: A review process designed to reconstruct and inspect decisions after they occur.
Calibration: The process of aligning reviewers, rubrics, or model confidence with observed reality.
Evidence packet: The structured context a reviewer needs to judge an AI output or action.
LOOP-SAFE: The book's oversight framework: Locate decision, Order by risk, Observe evidence, Price loop, Sample intelligently, Align reviewers, Feed system, Escalate and exit.
Post-action sampling: Reviewing a subset of outputs after the system acts.
Pre-action review: Human review required before an AI action is executed.
Risk tiering: Routing outputs or actions based on harm, reversibility, uncertainty, novelty, and policy sensitivity.
Trace evaluation: Evaluating not only final output but the steps, tool calls, permissions, and side effects that produced it.
Key Takeaways
- Approval loop: A pre action human review process that must approve an AI action before execution.
- The practical test is whether a team can name the evidence, owner, and failure mode before it changes behavior.
- Read this with Human in the Loop Is Not a Plan and the adjacent chapters when you need the wider Evals and Evaluation frame.
Operational note
This short chapter is a map, not a full argument. Its job is to keep Appendix B: Glossary usable inside Human In The Loop: define the terms, point to the sources, and make the next decision easier to replay.
Use it as a checkpoint before you treat the surrounding chapters as advice. A reader should be able to name the claim, the evidence, the risk boundary, and the follow-up page without interviewing the original author. If that replay fails, the chapter has not done its job yet.