Grading answers for faithfulness to their sources with a frontier model costs $2.52 per thousand items and scales with your traffic. Ten small open-weight judges from unrelated families, combined by a plain logistic regression, cost about four cents per thousand and keep a median 93% of the frontier judge’s agreement with the gold labels once chance agreement is discounted. It is not a blanket win — across eight sets the panel clearly beat the frontier judge on one and clearly lost on three — and the transferable part is not the panel but the cheap test that tells you, before you switch, whether yours is one of the cases where it works.