From San Digital, who run live AI courses for engineers and business teams.Live AI training from San Digital. See the courses

declawd

Articles

Research notes from Declawd. Each one names its evidence and limits, with links to public sources. No score here should be used to judge a person.

The pixels change

Published ; updated

A registered Rust/Python comparison retains decoder differences, an unmarked positive and the cost of pixel removal.

When the null shifts

Published ; updated

Registered synthetic checks show how repeated evidence and a small null mismatch can break count and corpus bounds.

Choose the right check

Published ; updated

Provider watermark and file-check routes, their access limits and the local tools Declawd supplies.

Every window

Published ; updated

Exact enumeration of 4,096 synthetic sequences measures false alarms across three window families, with every uncorrected and corrected outcome retained.

A third registration

Published ; updated

A new registration fixes fixture and output checks. Its single marked fixture still missed the selected cut-off, and every result is retained.

November evidence

Published ; updated

New research, certificate checks and standards, with provider detector access and gaps in source captures recorded.

More windows

Published ; updated

Searching more windows raised human false positives. A separate cut-off reduced them while missing most marked insertions in the registered experiment.

Two carriers

Published ; updated

The matching TrustMark decoder recovered ten of twelve payloads before and after C2PA removal. Every cleaned file retained the marked image bytes.

A second registration

Published ; updated

Declawd v2 reviews every candidate in context and measures the registered 200-pair minimum, while preserving v1 evidence.

What public cleaners change

Published ; updated

A public cleaner changes what it can see: visible Unicode, metadata, wording. Is Claude's text mark gone afterwards? No cleaner can tell you that without Anthropic's matching detector.

Six word changes against our published watermark

Published ; updated

It took five substitutions to carry the fixed passage across Declawd's published cut-off. The sixth was insurance, worth more than two scoring steps of headroom.

A byte-level scan of collected Claude output

Published ; updated

120 files, four Claude model assignments, and along the way a test of which controls survived collection. One caveat up front: a scan like this cannot see a word-choice watermark.

Em dashes and curly marks in two corpora

Published ; updated

More em dashes and curly marks in five novels than in 120 collected Claude files. And nearly every em dash Claude did produce sits in dialogue.

One article, three sets of style instructions

Published ; updated

Tighter instructions and a checking pass shifted five familiar style patterns. Reproducing the result is harder, because the rubric is unpublished.