100%
published fixtures passing
Public proof, version 1.0
See exactly what changes, and what does not.
These downloadable fixtures run against the same scanner used by the website. Every result is deterministic and independently reproducible. This benchmark does not test statistical model watermarks or claim AI authorship detection.
13/13
expected results matched
0
visible characters rewritten
| Published fixture | Result | Exact input → output | Behavior |
|---|---|---|---|
| Zero-width space U+200B | Pass | "watermark"→ "watermark" | Removed deterministically; visible words unchanged |
| Word joiner U+2060 | Pass | "cleartext"→ "cleartext" | Removed deterministically; visible words unchanged |
| Right-to-left override U+202E | Pass | "safetext"→ "safetext" | Removed deterministically; visible words unchanged |
| Unicode tag U+E0061 | Pass | "plain text"→ "plain text" | Removed deterministically; visible words unchanged |
| Variation selector U+FE0F | Pass | "text️"→ "text" | Removed deterministically; visible words unchanged |
| Hangul filler U+3164 | Pass | "twoㅤwords"→ "two words" | Removed deterministically; visible words unchanged |
| Halfwidth Hangul filler U+FFA0 outside Hangul context | Pass | "twoﾠwords"→ "two words" | Removed deterministically; visible words unchanged |
| Mongolian variation selector U+180F outside Mongolian context | Pass | "clear᠏text"→ "cleartext" | Removed deterministically; visible words unchanged |
| Soft hyphen U+00AD | Pass | "watermark"→ "watermark" | Removed deterministically; visible words unchanged |
| Byte-order mark U+FEFF | Pass | "cleantext"→ "cleantext" | Removed deterministically; visible words unchanged |
| Non-breaking space U+00A0 | Pass | "two words"→ "two words" | Removed deterministically; visible words unchanged |
| Bell control U+0007 | Pass | "safe\u0007text"→ "safetext" | Removed deterministically; visible words unchanged |
| Cyrillic lookalike U+0430 | Pass | "pаypal"→ "pаypal" | Reported and preserved |
Language eligibility and exclusions
Literal code-point rules are language-agnostic, but context-sensitive characters are not automatically removable. Mixed-script lookalikes are reported and preserved. Mongolian selectors and Hangul fillers are removable only in fixture-declared out-of-language contexts. This benchmark publishes deterministic coverage, not a fixed false-positive rate; no representative multilingual corpus has yet established one.
What this proves
- Exact published inputs produce the displayed outputs.
- Supported removable characters are cleaned without rewriting readable words.
- Suspicious visible lookalikes are reported but preserved.
What this cannot prove
- Whether an AI system authored or processed the text.
- Whether a provider-level statistical watermark exists.
- Whether an unavailable provider verifier would return a match.
Reproduce the result
Download the versioned JSON fixtures, inspect the open-source scanner, and run the automated tests. The release gate fails if the downloadable fixtures and tested source ever differ.
Compare lossless cleaning with rewriting