G'day andreas1234567,

Just a thought that may provide a partial solution. I haven't tried this myself and I don't know applicable it might be to your situation.

Several assumptions are implied, including: existence of reference documents, lack of variable content (e.g. document-specific reference/ID numbers, date/time fields, etc.), and so on.

If you took a Digest::* of reference documents, you could compare with digests of test-generated documents.

While this won't identify specific issues, it might reduce the "pile of documents" to a "handful of documents" that require closer, subsequent scrutiny (whether by a manual or automated process).