Run the same documents through models from cheapest to frontier and measure, on your own held-out set, what each step of the ladder is worth. Then ship: your pipeline graded against labels you have never seen, rerun twice with the database diffed, killed mid-run and resumed. Your harness ships with the pipeline. You also write your first solution memo: one page, for a reader who does not code.
It ends with
Precision, recall and cost per document verified by the harness, a routing decision you can defend with measurements, and the memo.