Content Provenance Hygiene

Inspect first, change only authorized provenance/metadata surfaces, and verify the result without making authorship claims.

Trigger boundary

Use this skill for authorized inspection or sanitation of invisible Unicode, C2PA/Content Credentials, EXIF/XMP, PDF/document properties, or similar provenance/metadata surfaces.

Do not trigger for detector evasion, authorship misrepresentation, unrelated rewriting/image editing, access-control bypass, or generic file transformation.

Required inputs

Resolve the exact artifact/scope, ownership or authorization basis, inspect-only versus clean intent, visible-content preservation invariants, approved metadata classes, output expectations, working-tree/file state, and which inspection tools are available locally.

Safety baseline

• Preserve the original artifact by default.
• Inspect before mutation.
• Never claim metadata removal proves human authorship or defeats detection.
• Do not silently install tools, start services, or upload artifacts to a remote cleaner; uploading discloses the content.
• Treat unsupported container capabilities as unknown, not success.

Workflow

1. Freeze scope and intent. Record artifact identity, preservation invariants, and allowed metadata classes.

2. Resolve tooling. Read references/inspection-toolchain.md, probe which inspectors are available, and mark surfaces without one as not inspectable. For text, run the read-only scanner:

   python skills/content-provenance-hygiene/scripts/inspect_text_invisibles.py PATH... --format markdown

   It reports zero-width, bidi-control, tag, variation-selector, soft-hyphen, and other format characters with file:line:column, never modifies files, and exits 1 when it finds anything.

3. Inspect. Run the available inspectors read-only and classify each surface as confirmed, probable, informational, not detected, or not inspectable.
4. Plan the smallest remediation. Use references/remediation-boundaries.md; separate deterministic metadata cleanup from substantive content edits.
5. Clean only approved scope. Preserve visible pixels/text/structure unless the user explicitly requested otherwise.
6. Re-inspect. Compare target findings before/after and verify output integrity against the preservation invariants.
7. Report. State what was found, what changed, what could not be inspected, and what the result does not prove.

Handoffs and interaction boundaries

Substantive writing/image/document edits belong to their owning workflow. Cleaning-tool failures with an unknown cause go to debugging-investigator. Product metadata behavior may involve documentation synchronization. Final release claims belong to verification-and-release.

Failure handling

If no available tool can inspect a requested surface, report the unsupported capability and stop before destructive conversion. If cleanup changes visible content unexpectedly, treat that as failure and preserve the original.

Stop conditions

Stop when the authorized metadata scope is inspected, approved cleanup is verified against the original, integrity is checked, and unsupported surfaces or limitations are explicit.