Merci, Johan, de tes commentaires.
Unfortunately, I only have a model and no hard evidence for the claim that refactoring reduces volatility in the cost of new features. Measuring this stuff creates all kinds of problems, because of the sheer number of confounding factors in the research. I also have only an intuition that the cost of refactoring doesn't dominate the costs due to unhealthy design. I think the difference shows up most in support costs, but you can't measure support costs that you don't incur. :) I've seen organisations spend 70% of their development budget fixing problems, and I think that we can design ourselves away from that problem, but I haven't got the permission to run a controled-ish study-ish with that hypothesis.
Not only that, but we have to be careful about measuring the learning curve, just one of the many confounding issues that makes researching this topic difficult.
Please, please, please share your results with us as you go.