I'm slowly but surely working through my holiday backlog of email and weblogs. Slowly being the operative word here.
Anyway, Stuart has a great post on the process by which we build models. And he's not talking theoretically here, he's working on a model for the designer definition file in the DSL toolkit. (Which is good news in and of itself as hand-writing the XML dsldd file is a pain in the butt. Though until then there's the great Dm2Dd tool from Modelisoft). The iterative process he describes certainly looks a lot like development, in the same way that C# development looks like C development. Similar steps taken on different concepts. Additionally, he's working bottom up - the output of the model will eventually be a working program (a designer in this case) which is the point I made in Abstraction Gap Leapfrog. There are existing abstractions that work now (i.e. the code generated from the existing dsldd file) and he's trying to building something one level up from there.
I also like Stuart's use of "fidelity" instead of my use of "complete". Stuart uses it as an indication of how correct a given model is. That's what I was implying when I said "complete" but "fidelity" captures the idea much better. I could imagine both lo-fidelity and hi-fidelity models for a given domain, though I would imagine you would always want to use the highest fidelity model available. The difference would be the amount of custom code you have to write - the higher the model fidelity, the less code you write by hand. And I would imagine the model's fidelity would evolve over time, which introduces interesting questions regarding language evolution as well as the evolution of projects built with those languages.
I hope Stuart keeps blogging about this project.