Add verify-watermark-removal skill

[Yurakonoplya]

Summary

Adds verify-watermark-removal, a skill that measures whether a "watermark remover" or "humanizer" actually removed a statistical text watermark, instead of repeating the tool's own claim.

The check is possible because the mark is planted first, under a key the user holds. A detector that knows the key is the strongest detector that can exist for that text, so the score is a ceiling rather than a guess. The skill also reports what the rewrite cost: meaning kept, facts kept, longest verbatim run, share of words changed, length ratio.

It only measures. It removes nothing, and it never claims anything about a specific vendor's watermark.

Component Details

• Name: verify-watermark-removal
• Type: Skill
• Category: testing-qa
• Backing tool: https://github.com/Yurakonoplya/unmark-checker (MIT, Python library and CLI, runs on CPU)

Testing

• Ran validation (node scripts/validate-skills.js and node scripts/validate-all.js, both pass)
• Tested functionality: the four CLI commands the skill drives (generate, check, check --json, matrix) are the ones the tool ships, and the example output in the skill is the real output format of unmark-checker check
• No overlap with existing components: no skill here plants or scores a text watermark

Examples

1. "I ran this paragraph through a humanizer that promises to strip AI watermarks. Did it work?" The skill generates a keyed sample, asks the user to run the tool on it, scores what came back, and answers mark_present, uncertain or mark_gone with the cost columns.
2. "Which of these three removal services is worth paying for?" Each run is recorded as a service file, unmark-checker matrix turns the folder into one table, and the services are compared on the same measurements.
3. "Is my text undetectable now?" The skill refuses that framing: clearing this mark has not been shown to clear any vendor's mark, and it says so instead of rounding a passing run into a promise.

[davepoon]

Thanks for the careful write-up — reviewed for security, prompt injection, and fit with this collection.

Security / prompt injection: The skill markdown is clean (no override/exfil instructions). It only measures; it never removes a mark and refuses the “now undetectable” framing. Key handling is sensible (read -rs, keep it out of files and shell history). Spot-checked unmark-checker at tag v0.1.5: local transformers/torch stack, no outbound HTTP client in the detector/CLI path. Pinning the git tag instead of main is the right call.

Codebase fit: Frontmatter validates (name matches the directory, category: testing-qa, license allowed). No overlap with an existing skill. External CLI dependency is documented the same way other install-from-clone skills are.

Merging.