I don't see what's wrong with including language features as extensions rather than bundling them into the core language the way Scala does. Sure, you have to e

I don't see what's wrong with including language features as extensions rather than bundling them into the core language the way Scala does. Sure, you have to explicitly turn them on in your file, but that's trivial. In practice, it's really no different from importing a library. At the very least, the GADT approach is not hard in Haskell. I don't see how it's any more difficult than the Scala approach. If anything, it's more natural: this is exactly the sort of use case GADTs were designed for. The main advantage of using GADTs here is that they're simple and explicit. Sure, you could do the same thing with case classes, but it would be more awkward, and you'd really be rebuilding part of the same functionality yourself. You don't really have to use data kinds in my example, but it makes the code nicer because you explicitly limit the types you can use as flags to 'Checked and 'Unchecked. You could have just made Checked and Unchecked empty types and stayed with normal kinds; the data kind approach is simply more elegant. The GADT method also ensures that the pattern matching is always safe. It also preserves exhaustiveness checking, which is very nice. I'm not sure if the Scala approach can do this. Another advantage over using witness objects (assuming I understood your idea correctly) is that you do not introduce anything new at the value level. 'Checked and 'Unchecked are only ever types, so the actual AST ends up being exactly the same as before just with a different type. I think this is a boon to clarity and simplicity. It also means that the use-site of the AST type doesn't change very much, if at all. Really, the main idea is that this is very easy in Haskell, basically using a single simple abstraction: the GADT. Just being hidden behind a language extension does not magically make the concept any more complicated or difficult to use! As an aside, if you want polymorphic variants, I don't see any reason not to use OCaml. It's a very nice language, and gets sub-typing right. All without sacrificing type inference.