Text watermark notes
Deterministic marks
Invisible Unicode, bidirectional controls, tag characters, exotic spaces, and selected confusables can carry machine-readable signals or cause broken copy, search, and diffs.
clean_text.py removes or normalizes known carriers and reports exact counts. By default it preserves contextual characters used by emoji, joining scripts, Mongolian selectors, Khmer inherent vowels, Hangul fillers, and Arabic/Syriac orthography. The lightweight workflow also disables space normalization by default so multilingual typography remains intact. Aggressive flags can damage intentional text and should remain opt-in.
Statistical marks
Token-sampling watermarks such as green-list or tournament-sampling schemes live in word choice and token sequences rather than metadata. A substantial rewrite can weaken such signals by changing syntax and vocabulary.
This is best-effort:
- no bundled detector has the vendor's secret key
- a rewrite generated by the same provider may introduce a new signal
- short or predictable text provides little statistical evidence either way
- detector behavior can change independently of this skill
Never describe a successful rewrite as certified, undetectable, or proof of human authorship.