I ran more than ten manual audits to make one tool respond well.

That worked once. It does not scale.

If a tool depends on judgment, the repeatable part of that judgment has to become an evaluation layer. Otherwise the author becomes the test suite, and every improvement quietly reopens old failures.