What is instruction provenance?

Definition

Instruction provenance distinguishes trusted system policy, owner commands, approved skills, tool output, retrieved content, and untrusted page text. It can record source identity, version, signature, retrieval path, timestamp, transformation, and the permissions under which the instruction may act.

Provenance helps prevent external content from silently becoming authority. Agents and execution systems can reject, isolate, or request review when an instruction comes from a mutable or untrusted source that lacks permission to direct consequential actions.

ELI5

Instruction provenance is the history of where an AI instruction came from and why the system should trust it. It separates real rules from text that merely looks like a command.

For example, a project owner's approved instruction can carry authority, while a sentence found on a webpage should remain untrusted content. Recording the source helps the system avoid letting the webpage control private tools.

Frequently asked questions

What can an instruction provenance record contain?

It can contain source, owner, version, signature, hash, retrieval path, timestamp, transformations, trust level, scope, and allowed authority.

How does instruction provenance reduce injection risk?

It lets systems distinguish untrusted retrieved text from authorized instructions and block authority that the source does not possess.

Videos explaining instruction provenance

  1. Portrait of Nate B. Jones beside the words Agents Need Boundaries