Illustration: the pirate captain writes in the leather-bound captain's log while a small ghost holds up a tally slate

Captain's Log — MarkMyAss Benchmarks#

Every supported test. Every failure. Nothing hidden.#

This page is the ship's log: generated at server startup by actually running MarkMyAss's public, reproducible test corpus through the real inspect → clean → verify pipeline -- the same one you get from ghostmark inspect / ghostmark clean / ghostmark verify. Nothing on this page is a hand-typed number. Failures, if any, are shown here, not hidden.

← Back to the MarkMyAss cleaner · AI Watermark Lab →

Summary#

GhostMark 0.5.0 · 4 fixtures · generated 2026-08-14T13:11:48+00:00

Detection: 4 / 4 matched documented expectations
Cleaning: 4 / 4 matched documented expectations
Independent verification (ExifTool 13.25): 3 / 3 applicable fixtures confirmed clean

0 known failures in the current corpus.

Per-fixture results#

Fixture	Detected as documented	Cleaned as documented	Independently verified
`text/hidden-unicode.txt`	✓ Yes	✓ Yes	N/A
`jpeg/exif-xmp-iptc-comment.jpg`	✓ Yes	✓ Yes	✓ Yes
`png/exif-xmp-text.png`	✓ Yes	✓ Yes	✓ Yes
`pdf/docinfo-xmp.pdf`	✓ Yes	✓ Yes	✓ Yes

What this corpus does and doesn't cover#

The corpus currently covers hidden Unicode (text), EXIF/XMP/IPTC (JPEG), EXIF/XMP/PNG-text (PNG), and DocInfo/XMP (PDF) -- the mechanisms MarkMyAss's support matrix lists as Verified. It does not currently include synthetic C2PA/JUMBF fixtures, so this page makes no pass/fail claim about C2PA -- that mechanism's status remains what /lab/c2pa says it is: Partial, regardless of what this corpus does or doesn't test. Statistical/model-level text watermarks are not benchmarked here either, for the same reason they're not benchmarked anywhere: no public detector exists to test against. See /lab for the full, honest capability matrix.

Reproduce this yourself#

git clone https://github.com/bens777/MarkMyAss.git
cd ghostmark
pip install -e ".[dev]"
pytest tests/test_corpus.py -v

Every row on this page corresponds to an assertion in tests/test_corpus.py, run against the fixtures in src/ghostmark/corpus/ per the expectations documented in manifest.json. Machine-readable capability data (not this specific benchmark run) is also available at /api/lab/status.

Something outdated or inaccurate?#

If a fixture's documented expectation looks wrong, or you think the corpus should cover something it doesn't yet, open an issue or submit a pull request against src/ghostmark/corpus/manifest.json and scripts/generate_corpus.py.

Last reviewed: 2026-08-13