Find supported invisible Unicode in your browser, remove only what you select, and optionally rewrite authorized text sentence by sentence. NoLLMWM never labels an AI-likelihood score as watermark proof.
We can identify a documented set of physical Unicode signals. Statistical or vendor watermarks cannot be confirmed without the vendor’s private verifier. Rewriting may alter those signals; it is not a guarantee.
Sent to the configured OpenRouter model only after you ask. Numbers, names, URLs, emails, code, and quoted passages are protected and checked after rewriting.
Requires an account and uses your character balance. Rewriting changes language and may reduce a statistical signal, but no removal outcome is guaranteed.
Your pasted text stays in the browser during the Unicode scan. The inspector names each supported control character, its code point, and its position. It separately reports preserved emoji joiners and complex-script formatting rather than deleting them indiscriminately.
Clearly unnecessary zero-width spaces, byte-order marks, and stray Unicode tags can be preselected. Soft hyphens, word joiners, bidirectional controls, and script-sensitive characters require your decision because they can carry legitimate typography or language behavior.
When you explicitly request a rewrite, the server protects numbers, names, URLs, email addresses, code, and quoted passages before calling the configured OpenRouter model. It then rejects altered placeholders and returns a compact integrity report for your review.
An invisible character is a literal code point in the text. A browser can enumerate it deterministically. A statistical text watermark instead changes how a model chooses tokens across a passage. The visible words are the signal; there may be no hidden character to delete. Vendor verification normally requires a secret key, an unpublished detector, or both.
NoLLMWM keeps these claims separate. “No supported physical signal found” means only that our local character list did not match. It does not mean “no watermark.” Likewise, a large rewrite can disturb token statistics, but any claim of guaranteed removal would be misleading without access to the actual vendor verifier.
Compare watermarks and AI detectorsThe inspector and selected cleanup run in your browser. An authorized rewrite sends the current text to our application server and then to the configured OpenRouter model, but the application does not save the source or rewritten text in Postgres. Usage records contain request IDs, character counts, model name, style, and outcome only.
Authentication uses a signed, secure, HttpOnly session cookie. Passwords use Node’s scrypt with a random salt. Rewrite characters are reserved transactionally before a provider call and refunded if the provider cannot safely return a result. Payment callbacks are accepted only after raw-body HMAC verification and idempotency checks.
Review the security modelNoLLMWM is for people editing their own drafts, teams processing content they are authorized to publish, and developers cleaning accidental control characters from text pipelines. It is not designed to misrepresent provenance, evade an academic or employment policy, remove evidence from someone else’s material, or support fraud.
You remain responsible for checking meaning, citations, names, quantities, quotations, and applicable disclosure rules. The automated integrity report is a guardrail, not editorial approval. If a publisher, school, client, or regulator requires AI-use disclosure, rewriting does not remove that obligation.