Meaning Audit Workflow v1.0 ships as a Claude Code Skill

The workflow's two modes matter for how far you can trust what it tells you. When a document arrives without its underlying sources, Meaning Audit runs Artifact-Only: it cannot judge fidelity to evidence, so it does not declare anything wrong, only that a claim's basis is not traceable inside the document itself. The author's honesty about that boundary — and about the missing origin of the test deck, whose creation method could not be confirmed from the document — is part of the design rather than an oversight.

A striking detail from the field test is what happened after the 26 findings came back. None were adopted as-is. Human review rewrote four and appended a note to one, in one case downgrading a verdict of "internally contradictory" to "not shown". The author describes the process as auditing the audit — the output itself goes through the same treatment.

The reported run-to-run variation is the practical caveat. Finding counts, finding types and even the main Document Class shift between runs, so this reads less like a repeatable inspection and more like a structured prompt for human judgment. The author's own closing note is modest: the clearest result so far is stopping to look at figures one would normally read straight past.