hierarchy

Definition

An explicit ranking of authority. In LLM safety, the instruction hierarchy orders sources by trust — platform or system message, then developer, then user, then tool output and retrieved content — and the model is trained so a lower tier cannot override a higher one, which is the main structural defence against prompt injection. It ranks trust in instructions, not the containment hierarchy of classes or network layers.

In the textbook

• The Memory Hierarchy and the Memory Wall