Editorial policy

Who writes this. One person, named on the about page, self-funded, with no investors, sponsors, advertising or affiliate commissions. Guides carry a visible author and an update date.

Sourcing. Claims about Anthropic’s watermark come from Anthropic’s own published documentation or from the peer-reviewed literature it points to, and are re-checked against the source on a schedule rather than written once and left. The model status table records the date it was last verified, and that date only moves when the check is actually run.

Claims about other people’s products. Quoted verbatim from their own pages, with the URL and the date it was read. We do not characterise a competitor’s tool from its reputation, and we do not describe what a product does internally unless we have read the code.

Numbers. We publish a figure only if we measured it or can link the measurement. If something is unverified it is labelled unverified. This rule exists because we broke it: see the corrections below.

What we refuse to claim. That we detect Anthropic’s statistical watermark, that we remove it, or that any text is “undetectable”. Anthropic holds the detection key and has not released it to anyone, so no third party — this site included — can verify whether a passage carries the mark or whether any edit removed it. Any product telling you otherwise is describing a check it cannot run.

Corrections. Errors are fixed on the page and recorded here rather than quietly edited away. If you find one, email ofiryieldchain@gmail.com.

15 August 2026 — an unsourced statistic. The guide on why watermark removers cannot work opened with “Five of the six tools ranking for ‘Claude watermark remover’ strip zero-width Unicode characters.” Nobody had counted. The figure was removed the day it was questioned and replaced with a claim the quotes support. A later audit found the original number was not merely uncounted but pointed the wrong way: of six results examined, two overclaim, three are careful about their limits, and one is mixed.

15 August 2026 — an overclaiming code comment. The sample text loaded by the checker was documented as carrying “one of every artifact class the engine detects”. It carries six of seven; it contains no exotic Unicode space, so that class never fires on it. The comment now says what is true.

16 August 2026 — model data was incomplete. Claude Mythos 5 had shipped and was missing from the model status table entirely, and Claude Fable 5’s launch date was an approximation where Anthropic publishes an exact one. Both fixed against Anthropic’s documentation. Claude Opus 5 and Sonnet 5 still have no published launch date, and the table no longer implies otherwise.