inspect-content-provenance

name	inspect-content-provenance
description	Inspect local files or text for observable provenance metadata, possible C2PA markers, hashes, and text control characters. Use when a user asks whether an asset contains Content Credentials, an AI mark, provenance metadata, or suspicious invisible characters and a cautious inventory is needed before cryptographic verification.

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

1. Preserve the source; never rewrite it.

2. For a file, run:

   python3 scripts/inspect_file.py /absolute/path/to/asset
   

   For 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.

3. Read manifest_presence as 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.

4. Check c2pa_markers[].confidence:

   • STRUCTURAL: found where a manifest belongs (PNG caBX, JPEG APP11 JUMBF, BMFF uuid, WebP C2PA).
   • MODERATE: a format-appropriate key or namespace reference.
   • SIDECAR: a detached .c2pa manifest beside the asset.

5. c2pa_mentions is separate and is never evidence. A document that discusses Content Credentials is not a signed asset and must not be routed to the verifier.

6. If any marker is present, use verify-content-credentials.

7. 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-watermark to analyse them.
• Absence of metadata is not evidence of human creation.

Read references/evidence-model.md when explaining ambiguous or unknown states.