Transparent by design
How the checker reaches a result
Every result is separated by evidence type. Literal characters are deterministic. Provider conclusions require an official verifier. The product does not infer authorship from formatting.
Literal Unicode
Deterministic
Exact code-point matching for zero-width characters, direction controls, Unicode tags, special spaces, control codes, and related formatting artifacts.
Visible lookalikes
Review only
Mixed-script characters are reported for manual review. The cleaner does not rewrite visible letters or guess the intended character.
Experimental rewrite proxy
Opt-in research
Local variants are scored with a reproducible open proxy. Short text remains inconclusive, the original is preserved, and a lower score is never presented as provider verification or proof of authorship.
Provider watermarks
Adapter required
A provider conclusion appears only when an official supported verifier is connected. An unavailable result is never reported as not detected. Even a verified watermark signal is supporting evidence, not proof that the generation pipeline, watermark key, or entropy source was uncompromised, and it does not by itself prove authorship.
Proof Receipt v2
Privacy-safe evidence
SHA-256 fingerprints bind the receipt to exact original and cleaned text without embedding either text. Findings, capability states, versions, and limitations remain inspectable.
File provenance
Local reference inspection
Format-aware header detection covers PNG, JPEG, WebP, AVIF, HEIC, SVG, PDF, and document containers. C2PA references, EXIF, XMP, properties, and embedded data are reported separately and never stripped. A reference is not signature verification.
Reproducible core
The literal scanner is available as a dependency-free MIT-licensed package with tests, TypeScript types, and a CLI. Review it on GitHub or install it from npm.