Fine-tuned 7B beats frontier on banking classification

An internal deck in October claimed that a LoRA-tuned 7B open-weight model on 12k labelled banking transaction descriptions outperformed a frontier model on a 14-class intent task, at a fortieth of the per-call cost. The number was real. The comparison was not: the frontier baseline had been run zero-shot with a one-line prompt and no examples, and nobody had checked whether that was the strongest reasonable configuration.

A three-day re-run held the eval fixed and gave the frontier model the same class definitions the tuning set encoded, plus twelve examples. The 7B scored 0.86 macro-F1; the frontier baseline moved from 0.81 to 0.92. The cost gap remained real but the accuracy claim reversed. Distillation from the strong baseline, which the original plan had skipped, closed the gap to 2 points and is the path r-slm-edge now describes.

The refuted claim is the narrow one in the title. Small models still win the edge and the cost argument; x-slm-edge-classifier is measuring exactly that on an in-store workload. What is dead is any comparison where the frontier baseline is not measured on the same eval with the same care as the tuned model.