Repository navigation
Ask the writer about negated setups, one at a time (1.1.0) - #3
Merged
Merged
Conversation
Stock phrases flag human posts or sit at chance; length tracks code and headings, not prose; the negated-setup check (A1c) raised the judges' mean P(AI) on rewrites from 43.6 to 46.4 across 20 titles, so it stays out. The numbers are in references/numbers.md so none of the three is proposed again without new evidence.
audit.py --review lists short "X isn't Y." setups with the reasons a reader may take them for AI. The rewrite loop never acts on them: fixing them there made rewrites read more AI to blind judges. After check.sh converges, and only with the writer in the conversation, the skill asks about each one in turn: keep, use the suggestion, delete, or write my own. The checklist, T1 and check.sh are unchanged.
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
A reader on Reddit stopped at "What surprised me most wasn't a number." in a post written with the skill. The checklist had not flagged it: A1 only matches when it, this or that is on both sides of the contrast.
What changes
audit.py --review [--json] FILElists negated setups: a sentence of 12 words or fewer that only denies something, opening a paragraph or right before the point. Each comes with the reasons a reader may take it for AI. It is not part of the checklist, T1 orcheck.sh, and it always exits 0.SKILL.mdgets a "Review with the writer" step. Aftercheck.shconverges, and only with the writer in the conversation, the skill asks about one sentence at a time: keep, use the suggestion, delete, or write my own. The rewrite loop is told to leave these sentences alone.references/numbers.mdrecords three checks that were measured and not shipped: stock phrases, length as verbosity, and A1c as a checklist item.CITATION.cffand.claude-plugin/marketplace.json.Why review only
As a checklist TELL, A1c looked right: 7% of held-out human posts against 31% of assistant posts, with held-out humans still at 0% T1 FAIL. End to end it made rewrites worse. Over 20 titles (8 fresh), 52 rewrites per version and two blind Opus judges on each, mean P(AI) went from 43.6 to 46.4 (Wilcoxon p = 0.036). Many of the sentences it removed were ordinary ones from the author's notes.
Measured on this branch
evaluate.pyon the held-out split gives output identical to 1.0.0.Not tested: the interactive question flow, which needs a person answering.