This is how I think about it. I care about asymptotics. What initial conditions (model(s) x workflow/harness x input text artefacts) causes convergence to the best steady state? The number of lines of code doesn't have to grow, it could also shrink. It's about the best output.