Inspect Content Provenance
Inventory evidence without making an authorship verdict. Treat located markers as clues and keep Anthropic text watermark status UNVERIFIABLE until Anthropic publishes a supported detector.
All paths below are relative to this skill's directory. If you are running from
elsewhere, use an absolute path to scripts/inspect_file.py.
Workflow
Preserve the source; never rewrite it.
For a file, run:
python3 scripts/inspect_file.py /absolute/path/to/assetFor literal text, run:
python3 scripts/inspect_file.py --text 'text to inspect'Literal input has no filename or host-language context, so this route checks the A.8 unstructured-text wrapper only. Use a file path when checking an A.9 comment or front-matter carrier.
Read
manifest_presenceas an inventory result:POSSIBLE: a structural carrier, sidecar, or format-appropriate malformed hint was located; inspect marker confidence and verify before relying on it.ABSENT: the bounded local scan found no structure; this does not prove human authorship.UNKNOWN: the input could not be meaningfully inspected, or the container is unsupported.
Check
c2pa_markers[].confidence:STRUCTURAL: found where a manifest belongs (PNGcaBX, JPEG APP11 JUMBF, BMFFuuid, WebPC2PA).MODERATE: a format-appropriate key or namespace reference.SIDECAR: a detached.c2pamanifest beside the asset.
c2pa_mentionsis separate and is never evidence. A document that discusses Content Credentials is not a signed asset and must not be routed to the verifier.If any marker is present, use
verify-content-credentials.Report the SHA-256 hash so later derivatives can be compared.
Non-negotiable language
- Say “possible C2PA marker,” never “valid credential,” unless a conforming verifier validates it.
- Say “Anthropic text watermark detection is unavailable,” never “no watermark found.”
- Invisible Unicode characters are a text-integrity signal, not an Anthropic watermark detector. Use
detect-text-watermarkto analyse them. - Absence of metadata is not evidence of human creation.
Read references/evidence-model.md when explaining ambiguous or unknown states.