Why did we build this?
This began as a Python research workbench for invisible characters and controlled watermark experiments. The public site adds browser tools for three questions: which writing patterns a passage contains, which Unicode characters came along with it, and which wording edits might improve a draft.
The detector and paraphraser now use algorithms written for Zyphh, and they run locally instead of sending text to an outside classifier or generation service. Their rules and limits are on the methodology page, and the original research lab still works on its own.
What do we care about in a result?
An observed character, a writing pattern and a verified watermark are different kinds of evidence. The interface keeps those meanings distinct. A familiar phrase cannot identify a model, and a signal score cannot reconstruct how something was written.
The detector reports learned signals with observable text measurements. The rewriter proposes bounded edits you can reverse individually. Both give you material to review rather than an unsupported authorship verdict.
How do we handle privacy?
Style analysis, paraphrasing and Unicode cleanup run in your browser. Your text isn't part of any page-asset or configuration request. We don't store documents in localStorage, publish result pages or keep a document-history database.
The optional statistical-watermark verifier is a separate integration that asks for explicit processing consent. Advertising, when enabled, stays on educational guide pages and away from text inputs and results. The privacy page covers hosting metadata and optional integrations.
What don't we claim?
We publish the detector’s measured benchmark results with the dataset, split method and limitations. Those results are not a universal accuracy promise or a comparison with competing products. No score reconstructs an author’s writing process, and no rewrite guarantees a detector outcome.
A result is useful when you can inspect it and decide what to check next. Our guide to AI detector accuracy explains how sample length, genre and the proportion of AI text in a collection change the meaning of a flag.
How can you send feedback or corrections?
If you spot an error in a guide or a tool result that looks wrong, use the contact page. Tell us which tool, any error code you saw and enough context to reproduce it, without sharing confidential material. A correction backed by a source helps most when a provider changes its documentation.