map-provenance-survival

Map Provenance Survival

Measure a real pipeline instead of assuming that metadata survives it.

All paths below are relative to this skill's directory. If you are running from elsewhere, use an absolute path to scripts/map_survival.py.

Workflow

1. Keep an untouched original and record every transformation.

2. Obtain each derivative through the actual user-visible workflow.

3. Run:

   python3 scripts/map_survival.py \
     --original original.png \
     --derivative resized=resized.png \
     --derivative cms-download=downloaded.png
   

For a directory of derivatives and a shareable report:

python3 scripts/map_survival.py \
  --original original.png \
  --derivatives-dir transformed-copies/ \
  --report survival-report.html

1. Add --c2patool /path/to/c2patool for cryptographic validation. Remote-manifest fetching is disabled unless --allow-network is explicitly supplied.
2. Treat LOST_OR_UNAVAILABLE as an observed pipeline outcome, not proof of deliberate stripping.
3. Read summary for the counts, then quote individual rows for detail.
4. Re-run after pipeline, CDN, export, or platform changes.

Shareable reports

• --report accepts a new .html, .htm, .md, or .markdown path and refuses to overwrite it.
• HTML reports are self-contained and script-free. Markdown reports have the same evidence and caveats.
• Absolute paths and the reproducibility command are redacted by default. --include-paths is an explicit opt-in.
• JSON always remains on stdout, even when a report is written.
• A report evaluates C2PA provenance survival only. It must not imply that proprietary pixel, audio, video, or keyed text watermark detectors ran.
• Directory batches skip hidden entries and symlinks, recursively sort files, and use relative labels. With multiple directories, labels are prefixed by the directory name.

Evaluating a transformation tool

This skill also measures what a metadata-removal or optimisation tool actually removes. Run the tool over a signed original, pass its output as a derivative, and report the observed matrix. Report only what was observed: this skill does not certify that any downstream detector will or will not fire.

Experimental discipline

• Change one transformation per derivative where possible.
• Hash every artifact.
• Keep content comparison separate from provenance comparison.
• Record the app, command, platform, and date outside the generated matrix.
• A different hash alone does not mean provenance failed; a derivative is expected to differ.
• Use a conforming verifier before describing evidence as PRESENT, VALID, or INVALID.
• Evaluate signer trust separately with verify-content-credentials; this survival workflow does not select a trust policy.

Read references/survival-testing.md for the state model and test design.