AI-driven audit of the item bank. Live mode polices items currently in the pool — flagged items drop to needs_review = 1 until human review. Pending review mode vets items that entered at needs_review = 1 (e.g. AI-generated content) — passed items promote to needs_review = 0 and become test-samplable.
Not authorized
This page is restricted to owner and md roles.
How this runs. Each batch calls Sonnet once per item with a strict consistency rubric. The mode controls the gate direction: Live pulls items at needs_review = 0 and demotes flagged ones; Pending review pulls items at needs_review = 1 and promotes passed ones. Every verdict writes an audit_log row, so already-audited items are excluded on the next batch and the loop converges. Refresh-safe: closing this page and reopening resumes from where it left off. Dry run runs one batch only — without DB writes the next batch would re-pull the same items.