A strong answer. “I would test operation rather than existence, with four things.

The erasure drill. Pick a synthetic subject and try to find and delete every copy. This exercises inventory, lineage, backups and exports simultaneously, and if it takes more than about an hour the inventory is fiction. I would run it as a drill rather than waiting for a real request.

Revoke rate on access reviews, not completion rate. A campaign closing at 98% completion with a 0.4% revoke rate is a rubber stamp, and completion rate is a vanity metric that rewards exactly that.

Classification granularity. Is it applied at column level or table level? Table-level classification means everything inherits the strictest tag, which means either over-restriction that pushes people into shadow copies, or a tag nobody honours.

A two-user test. Two accounts with different entitlements querying the same asset. If they see the same thing, the access model did not survive whatever pipeline produced that table — and this is where retrieval indexes for AI are currently failing badly.

Then the structural question: is governance a gate people pass through, or a default they get for free? A gate produces shadow pipelines. Classification applied at ingestion, access derived from it, retention scheduled automatically produces compliance nobody had to choose.”