What WordMark measures—and what it does not.

The product deliberately separates deterministic evidence from statistical heuristics. Hidden Unicode, control characters, metadata and protected-data preservation can often be checked exactly. Statistical authorship or proprietary watermark claims usually cannot.

Deterministic layer

The scanner enumerates Unicode code points, flags context-sensitive invisible characters, bidi controls, unusual spaces and mixed-script confusables. Safe Clean removes only artifacts with a deterministic cleanup rule; emoji joiners are preserved when used inside pictographic sequences.

Statistical laboratory

Metrics include token and sentence distributions, vocabulary size, lexical diversity, Shannon entropy, sentence-length variation, burstiness, repeated n-grams, average word length and punctuation density. These are descriptive features, not a binary “AI detector.”

Meaning and data preservation

Deep Clean masks code, URLs, dates, numbers, currencies, citations, named entities and technical tokens before transformation. A result is checked again afterward. The meaning score combines protected-data preservation with normalized content-word overlap. When a requested transformation falls below its configured preservation threshold, the engine automatically reduces the transformation strength.

Provider verification

The verifier adapter reports Unavailable unless the operator configures a compatible authoritative endpoint. The interface never converts “statistical transformation completed” into “watermark removed” without external verification.

Benchmarking

The repository includes deterministic tests and performance tests. Operators should add real-world fixture corpora for each supported language and document format before making production reliability claims beyond those tests.