Not model cards. Not marketing claims. We run fixed prompts through each model, keep the outputs, and measure what it costs — then publish the receipts.