Train/deploy a large local model as deputy — handles complex reasoning work cheaply. Slow is fine, quality is mandatory. Follows from Goal 4 benchmark results.