From 0.5/8 to 7/8 — and What That Number Hides
A controlled study on one desktop GPU: taking a local 30B model from 0/8 to 7/8 on a real code-review benchmark, one graded experiment at a time. Context, tools and checklists measured zero. Adversarial framing, model choice and pushed knowledge were the levers that moved. At local scale, the expertise lives in the harness, not the model. Continue Reading →