To ensure our software is secure there's a few tools we could build! We could monitor (using a reverse breadth-first search) which inputs ends a transcoding to ensure it matches expectations.
We can also have the Arithmetic Core feed it random input whilst running parsers on its output until we find some input which causes the output to become invalid. As that would indicate likely bugs. Running a breadth-first search afterwords could help find a simpler input that triggers the error!