The three-layer split is right, and Week 3 is the part most teams skip. Two additions from running seven agents on one property record. First, log the accept, edit or reject as a claim on the row, not as a metric on the side: who said it, what evidence they had, and bind the approval to a hash of the content so an edit after approval lapses it. Second, be careful with Week 4. Promoting a category to auto-send once acceptance passes 80% turns the human's stamp into the thing the model optimizes toward, and a reviewer stamping at throughput stops being a check and becomes ground. Keep a share of approvals issued against the underlying evidence rather than the draft, and keep the model's output at the lowest evidence grade no matter how often it was accepted. It cannot promote itself; a person promotes it. Your clean source-of-truth line is the whole thing. A source of truth is a record that says who may assert each value. If it does not say that, it is a source of values.