the human readable spec is tested by the standard battery of tests on a server/client
you still need the prose form, but this approach concentrates the implementation into one shared standard codebase which means additional features can be certified sooner
also, testing clients is more complicated you would have to create a client API spec as well for that, so that you can bundle your client code into blobs that get fed standard interface data inputs, ie forms