MG-SGLoRA is our attempt to keep adaptation cheap and reversible. In practice, this means carefully controlling which parts of the model can move and how we log those changes.
Early experiments show that naive adapters can silently drift core behaviour over time. A lot of the real work here is tooling: diffing behaviours, running targeted evals, and rolling back quickly when something looks wrong.
We'll publish more once the patterns stabilise. For now, this work is strictly internal and focused on safety and reproducibility.