Model comparison

Compare 2–4 generator and retriever configurations on the same dataset, judge and rubric. Pick a baseline, read the deltas, then promote the configuration you want gated for release.