assumptions

### Background

I wrote a configuration library in Python for my employer before discovering Confit. They share many of the same features, but I think there are some things mine does better than Confit. Instead of spawning a new project, I'd like to contribute to Confit, but not without finding a consensus on philosophy, features, and implementation. I'm submitting a few different proposals as separate Github issues.
### Assumptions

I want to start by discussing a few assumptions I used.
1. **A configuration consists of variable bindings.** In each new project, a configuration is born only when some variable holds a magic value (e.g. hostname of external service, port to bind, path to temporary directory, path to credentials). The magic value is moved to the configuration as a default, and the variable's value is loaded from the configuration. There is no need for structure richer than a mapping of names to values.
2. **Each variable has a type.** Equivalently, each variable can be validated.
3. **Each type can be serialized to and deserialized from a string.** Configuration files, environment variables, and command line arguments are just strings. We have to be able to represent values as strings.

Can we adopt these assumptions? Do we have real-world counterexamples that cannot be conveniently expressed with these assumptions in place? I'd like to hear your thoughts.