Question
What should a person be able to see before trusting an agent to continue?
A useful boundary
Context is not only memory. It is the boundary that tells a person what the system believes is in scope.
A readable trace
A short trail of sources and decisions is more useful than an exhaustive transcript.
Control without chores
The reader should be able to correct a premise without becoming the archive administrator.
Reflection
This remains a question set, not a claim of a solved interaction model.