See, when you are defining a Haskell program, you are conceptually creating a tree of thunks that eventually get executed by the GHC runtime. Those thunks are typed, meaning they have typed inputs and outputs, and are either side-effect-free or are defined in a context that controls their side effects (e.g. the IO monad).
So you can change the definition of types willy-nilly and either get a working compiled program or some error output that tells you exactly what is broken or doesn't make sense to GHC's model of your proposed computation. Because computation itself is typed, you have a stronger guarantee that it will work as expected when executed by the GHC runtime. Because side effects are controlled, you are forced by the type checker to handle runtime errors (or crash).
At least that's how I understand it as someone who works with gifted Haskell engineers, but exists very much on the periphery of understanding.
