Research

Open methodology before big claims. This hub will publish reproducible studies of observable artifacts in Claude outputs—not fabricated results about Claude’s keyed watermark.

Current studies

No corpus result is published yet. We are preparing a transparent study of zero-width characters, Unicode format controls, unusual whitespace, variation selectors, and supplied clipboard/file metadata across clearly labeled response types.

Required study record

Each release will include collection date, model/version, sample size, collection method, tests performed, raw findings, limitations, code, and update history.

Methodology and source code

The repository already includes deterministic text inspection utilities. The first public study will include its corpus permissions and reproducible scripts before any numerical claim is made.

Last tested: September 1, 2026