unmarkclaude.io

Checked service, our own processing

unmarkclaude.io review: our own processing on the common sample, 5 September 2026

Our own processing took the sample every other tool got, 992 words, on 5 September 2026. Mark not found at 0.8. 3 of 3 facts kept, 92% of the length, 74% of the meaning.

Address: https://unmarkclaude.io/

What came back, in five figures

Sample UM-D150B2, 992 words, 2026-09-05

The mark

0.8

Detector score, our key. Mark not found: found at 4.0 and above, grey band down to 2.0.

Meaning kept

74%

Sample and result cut into pieces and matched piece by piece, so moving paragraphs around costs nothing.

Facts kept

3 of 3

Numbers, dates, links, names and terms taken out of the sample and looked for in what came back.

Wording changed

91%

Share of the sample's words outside the blocks that stayed identical. Longest untouched run: 10 words.

Length

92%

Words back against words out. The halfway tick is the length of the sample itself.

Core pipeline with default settings: verbatim run limit 4 words, entity check on, lost entity falls back to the original sentence.

Its row among the Claude watermark removers .

Test us yourself. The same sample, our detector on what comes back. Take the marked sample