Anthropic official verifier
unavailable-2026-08-26unsupported
No official detector call is available to this page.
AI content toolsfrom Opace Public research boundary · Updated 26 August 2026
Claude Watermark Readiness Lab
Watch a watermark check pass with the right key and fail with the wrong one, on text we generated ourselves. It is the real maths, running in your browser. It cannot check anything Claude wrote, and nor can anyone else outside Anthropic.
Claude production verification: NOT AVAILABLE
SourcePublic fixture boundaryReproducible inputs only
SignalsProduction claim heldNo public Claude verifier exists
Watermark scanKnown keysReal SynthID mathematics; Anthropic production keys stay private
Try it yourself
Try the mathematics yourself
Most explainers about watermarking animate made-up numbers. This one does the real sum, in your browser, on text we generated ourselves and on anything you paste. Same published method, same arithmetic.
The demo detector is a 1.7 MB module (mostly the GPT-2 vocabulary). It loads once, on request, and everything then runs locally in this tab. Nothing you paste is sent anywhere.
Passage, coloured by per-token g-value
Each token shows its mean g-value across the six tournament layers. High values carry watermark evidence (green tint, solid overline), low values count against it (red tint, dotted underline), grey tokens are unscored (no full context window, or masked as repeated context).
Pick a sample to score it
Same text, wrong key: the signal vanishes. That is why a watermark can only be checked by whoever holds the key it was written with, and why no public verifier for Anthropic's production watermark exists today.
- Mean g
- –
- Weighted mean g
- –
- z-score
- –
- p-value
- –
- Scored positions
- –
- Key · depth
- –
Paste your own text
Local in this browser. Your text is tokenised and scored in this tab against all three demo keys, held in memory for this run only; it is not stored or sent to analytics or any server.
What editing does to the signal
The fixture set includes degraded copies of two watermarked passages, scored here under their own generation keys. Truncation keeps most of the per-token signal because the score needs no document structure, while word substitution replaces the very token choices that carry the watermark, so the mean falls towards 0.5; Google's published guidance says detection confidence can drop further under thorough paraphrase or translation.
| Passage | Tokens | Scored | Mean g | z | Outcome |
|---|
Dated status
Current Claude detector status
Anthropic's 14 August 2026 announcement describes watermarking for future supported Claude models and a detector planned for its API. No official production detector is available to this lab, so the canonical status remains unsupported with no score or threshold.
Public known-key demo detector
synthid-demo-v1 · lab 0.1.0attention
Runs on this page with public Opace demo keys and Apache-2.0 reference mathematics. Never evidence about Claude production text.
Protected-content check
core 0.0.0-privatenot run
The accepted local core can record exact factual anchors, but no fixture check ran on this page.
What rewriting changes
What rewriting changes
Google's SynthID guidance says detection confidence can fall after thorough rewriting or translation and that the method is not designed to withstand determined attacks. Signal movement and factual damage still need separate measurements.
Record the public signal
Name the scheme, key, configuration, fixture commit and detector version.
Measure factual damage
Check protected spans, edit distance and meaning separately from the signal.
Keep Claude claims unavailable
A public result cannot be relabelled as a Claude production result.
Required evidence
What would prove removal
A supported claim would need an official production method, the right key and configuration, a recorded source, a matched detector version and a valid before-and-after test. That evidence is not available, so this lab does not promise to remove a Claude watermark.
Reproduction
Reproduce the lab
The Google DeepMind SynthID Text reference repository provides a public known-key research implementation. It is not Anthropic's production verifier. The Claude SynthID text-watermark research note records the current claim boundary, while the AI content integrity methodology defines statuses, protected evidence and the acceptance rules for future fixtures.
The exact Opace repository and command-line reproduction links will appear only after their public destinations, licences and release artefacts have been approved.
Questions about Claude and public watermark tests
Frequently asked questions
Can this page verify a Claude production watermark?
No. Claude production verification is not available. Public research fixtures can test a documented configuration, but they cannot certify how an Anthropic production system behaves.
Does rewriting prove that a watermark was removed?
No. A changed public detector signal does not prove removal from Claude output. Rewriting can also damage names, figures, links, quotations and meaning, so those effects must be reported separately.
Why publish the lab before an official detector exists?
The lab makes the missing evidence visible and keeps the integration boundary ready for a future official method. It also gives researchers a clear route to the dated methodology without presenting a simulated result.
Where can I read the testing rules?
The AI Content Integrity methodology defines each result status, the protected-content rules and the evidence required before a public fixture can support a claim.
Created by Opace Digital Agency
Need help putting content checks to work?
Opace designs practical AI and publishing workflows for organisations in Birmingham, across the UK and internationally.
Dated research boundary
Public evidence is not Claude production verification.
Read the current research note before treating any public fixture result as evidence about a production system.