If you want to exhaustively describe the behavior of e.g. Rust 1.25.0, without a formalized model, then the only way to do so it to run it for all kinds of possible (edge-)cases you encounter and document the output.

If all you have as a reference is a bunch of test cases without any underlying reasoning as to why input A leads to output B, then that's not really a useful target to implement against. At that point you can just go the direct route, not bother with any standard and just implement directly against the Rust compiler.