November evidence
Published ; updated
The 2 November refresh keeps nine text-watermarked Claude models in the public coverage table and text detection in private preview. It adds newer research, certificate-validation developments and standards references. A publication, an implementation and a reproduced result each supply different evidence.
New papers
The TextSeal2 preprint was submitted on 29 October after our earlier review. Its source appeared in Meta's TextSeal repository the following day. The authors compare their method with their own textGrain implementation on the models and configurations in the paper. That comparison supplies no measurement of OpenAI's production watermark.
The paper retains important limits. Paraphrasing removed almost every tested mark. Its textGrain implementation slightly led under one diversity measure. The authors' unbatched, unoptimised implementation on one GPU preserved speculative acceptance but was about three times slower than plain decoding overall. Its localisation procedure counts a family of spans and applies a correction for that search. Our exact window-family experiment checks the correction in a finite independent model. The separate mixed-passage experiment illustrates why the complete search procedure needs calibration on its own data.
The review also identified soft-binding research published earlier in October. Its reported unmarked-image detector positives reinforce the distinction between a generic positive flag and recovering the expected payload. Our TrustMark experiment required the exact registered payload and retained both misses. Its twelve unmarked controls provide no population false-positive estimate.
Certificate checks
David Buchanan's 29 October report describes obtaining a Google Photos leaf signing certificate for an attacker-chosen key and questions revocation handling. Its demonstration signs a CMS message. It is not a reproduced Declawd image-validation test.
Upstream work now includes an option to prefer live OCSP checks over embedded responses and separate signer and timestamp-authority trust material. The conformance tool enabled the live-check preference on 2 November. A source change or merged PR establishes what that version implements. An older installed package needs its own check.
Declawd's inspector performs local checks with a pinned signer-list snapshot. Its receipt separates embedded-record presence, content binding, signature, issuer trust, timestamp trust and revocation. A chain matching that snapshot does not establish its current revocation status or the truth of the claimed creation history. Live revocation and remote-manifest requests remain unavailable in this offline policy.
The reported JavaScript settings problem was tested against our exact shipped package. The configured reader returned binding and signature checks, disabling verification removed those results, and malformed settings failed. The maintenance record records those controls and the shipped package hash. Those controls support the settings path used here. They establish no result for another package build or untested settings path.
Standards
C2PA 2.5 is the current specification reference in this review. Its text-wrapper detail and external-retrieval guidance need to be read separately from the media formats and behaviour supported by a particular SDK. The earlier four-state C2PA example retains its recorded versions and outputs.
The current Unicode notes use version 18 data for explanatory script and meaning information. The registered character classes, fixtures and earlier source pins remain intact. A newer annotation table does not change which transformations an earlier experiment measured.
Provider access
Anthropic's public model table remains unchanged in this capture. Its public file checker excludes text. OpenAI's current text-detector programme requires approval and excludes academic-integrity checking of student work. Google's public verification instructions cover images, audio and video. These separate access statements supply no interchangeable production text-detector key or threshold.
The EU marking task force has discussed a common detection API and lists later interoperability-group activity. The reviewed official minutes and task-force page establish discussion and meeting records. They supply no operational common detector or published interface in the material reviewed.
Search limits
The refreshed GitHub sweep starts at the same 15 August creation floor and includes 2,556 repositories. Eleven raw README recoveries were checked against their recorded sizes and Git blob hashes. One unidentified search match, nine unreadable heads and 22 README gaps remain. Known older projects and research indexes received separate reviews. Publication and revision dates were checked independently of crawl dates.
The source review, repository capture, repository review and research review record source identities, hashes, decisions and limitations. Third-party findings remain attributed to their authors, with reproductions identified separately.