Tell the week the watermark arrived, including the part that went wrong - #63
Merged
Conversation
The eighth article, in both languages. The news is days old — Claude has been marking its output since the second — so the timing is the best of the eight and the story is the least comfortable. It does not argue that the watermark leaves us untouched. It says I set out to prove that, could not, and found a worse fault in my own tool on the way: rewriting to strip a mark does not change whether we flag a passage (p = 0.453, hypothesis withdrawn), while cutting the same writing to four hundred words does — eleven of thirty documents flagged at one position in the text and not another. The reviewers are in it by name of finding: three independent readers, three false sentences of mine, one of them describing my own code wrongly. A project that has spent seven articles asking other people to publish their errors does not get to leave that out. The order of the three results is deliberate and should survive editing. A reader who leaves with "this tool catches AI-paraphrased work" has taken away the one thing the study could not show, so the negation leads and the social copy leads with it too. Both pass their own linter at 7/100 with burstiness 0.66. The English draft tripped our own em-dash rule at 19 in 1,775 words and was fixed before the HTML was generated, which is the same gesture the calibration article made. The cover is wordless apart from two scores, so one image serves both editions: one page, three windows, three verdicts, and the flagged stretch is the one whose lines run even. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015PEbbiYSNPw7jE3LrPNhyF
AI_Tasks/ is where the state of play sits between sessions: what is half finished, what only Pedro can do, briefs for the other models, and the committee's verdicts. Adapted from the ERP's folder of the same name, with one difference that governs the rest — that project has no issue tracker and this one does, so nothing here duplicates a GitHub issue. The test is whether a stranger could act on it: if they could, it is an issue. Committed as an ignore rule rather than left loose on purpose. An ignore that exists only in somebody's working copy is one `git add -A` away from putting unpublished drafts and unfixed defects on the public default branch, which has happened here once already with temp/. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015PEbbiYSNPw7jE3LrPNhyF
Owner
Author
|
Added one more commit on this branch: an ignore rule for It rides along here rather than in its own PR because leaving the rule uncommitted is the actual risk — an ignore that lives only in a working copy is one |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The eighth article, bilingual, plus its cover and the publication copy.
The angle
Not "the watermark doesn't threaten us". The article says I set out to prove that, could not, and found a worse fault in my own tool on the way. Three results, in this order:
p = 0.453. The hypothesis is withdrawn in bold.The order matters and should survive editing. A reader who leaves with "this tool catches AI-paraphrased work" has taken away the one thing the study could not show, so the negation leads — in the headline, in the article, and in the social copy in
PUBLICACION.md.Verified before generating the HTML
Both languages pass their own linter: ES 7/100 · burstiness 0.66 · ~1,850 words, EN 7/100 · burstiness 0.66 · ~1,780 words. The English draft tripped our own em-dash rule (19 in 1,775 words) and was rewritten to clear it, the same gesture the calibration article made.
Cover
social/watermark-story-cover.png, hand-authored SVG. Wordless apart from two scores, so one image serves both editions. It argues rather than decorates: one page of writing, three windows over it, three different verdicts — and the flagged stretch is the one whose line lengths run even, which is the mechanism visible without being named.Also opened alongside this work: #61 (five Spanish rules firing on ordinary formal Spanish) and #62 (the character scanner reading Turkish dotless ı as an artifact).
🤖 Generated with Claude Code