In summary: The more layers of interpretation your `eval` function needs to do (assumption of source language, tokenization/parsing rules, symbol definitions, etc), the less well-defined the semantics are.

So on the less-well-defined end, you have something like JavaScript's `eval` function, in the middle you have lisp macros, which get an AST instead of just source code, and on the actually-pretty-well-defined end, you have a lambda that's already been parsed/compiled and just needs to be evaluated.