JSON is a carefully-chosen subset of what "eval" will actually include, the three major differences being that it specifies the delimiters rigidly (JSON that uses apostrophe-delimited strings is not JSON), it rigidly specifies Unicode (defaulting to UTF-8), and it doesn't permit anything that looks like Javascript code. So, comparing it to "wiring over some Lisp and evaling" it is not accurate, JSON was created for the explicit purpose of not doing that. Previous approaches typically did.
That's what I do. I persist python objects into s-expressions using a c-extension module then parse and un-persist data back to python on the other side. Everyone asks me why I don't just switch to JSON. Well I might one day, I just like S-expression for now and the parser is very small, fast and was easy to implement. I even have references (like Yaml) to persist arbitrary object graphs.
