An answer can quote the policy correctly and still be wrong for the request. The missing detail is often the date the policy became effective.
Internal knowledge systems need three dates kept distinct: when a version was published, when it was approved, and when it takes effect. The newest file is not always the governing one.
For example, a policy approved in October takes effect November 1. A request dated October 25 and a request dated November 3 may need different answers. Each answer should name the version it relied on and that version's effective date. If the organization hasn't specified how in-progress requests transition, that decision belongs to the policy owner. The assistant must not invent it.
Build the evaluation as a small dated question set. Write questions on both sides of the effective date. Run them against the system and check two things: did it select the applicable version, and did it surface the effective date? Record which versions must remain available for historical questions, and assign an owner to retest whenever a policy changes.
This is an evaluation design, not a compliance certification. It does not guarantee that retrieval solves every conflict, and it does not replace legal review where the stakes require it.
Test one policy question on both sides of its effective date before expanding internal assistant access. The answer that ignores time is guessing.