Journal

Why we stopped fine-tuning and bet on RAG instead

After three fine-tuning runs and a Gemma rebase that regressed quality, we ratified ADR-007: the moat was never the weights. Here's what the benchmark data said and what we built instead.