The harness, not the model: how to make a weak or local model reliable enough to ship

If you've tried to point a local model — Qwen, a quantized Llama, whatever fits on your GPU — at a...