v0.1.1

Fixes from the first real pass on a short story and from the wiki run.

• Real corrections no longer count as AI-written words. A change between forms of one irregular word (says to said), between commonly confused words (road to rode, your to you're), or a one-for-one swap to a word the author wrote in the same or a neighbouring paragraph now counts as the author's. A pure addition never does. On the first real story the ledger went from 3 AI-written words, all false, to 0.
• Author text that copies a run of three or more words from a suggestion counts the common words in that run as AI-written too.
• ew.py suggest saves every valid suggestion, lists refused ones by item number and skips items already saved, so a resubmitted file adds only the fixed ones.
• The apply changelog says "the author wrote N new words" for author text. A refused intake (for example a PDF without --text) writes nothing.
• README Privacy lists the full works-store lookup order; kb/delivery.md names suggestions.json as the source of truth.

Tests: 31 unit tests on Linux, macOS and Windows (Python 3.9 and 3.14). Eval suite 9/9.

editwright.zip is the skill folder for claude.ai and Claude Desktop (Customize > Skills > Upload skill).

editwright 0.1.0

First release.

editwright edits writing like a professional editor and never writes on the page: editorial letters, developmental notes, line, copy and proofreading suggestions with IDs and quoted anchors, script coverage and nonfiction fact-check flags.

• The author's file is read-only: hashed at intake, checked at the end of every run.
• Suggestions, not rewrites; anything that needs new wording is a question for the author.
• A provenance ledger counts every AI-written word. Human-authored mode allows none in fiction and 1% (at most 50 words) in nonfiction; copying a suggestion's wording still counts as AI-written.
• Only the IDs the author accepts are applied, to a copy, with a changelog.
• ew.py: one standard-library Python CLI (3.9+) for intake, word-safe cleanup, counts, chunks, suggestions, the ledger, apply, a tracked-changes review.docx and author preferences.

Tests: 26 unit tests on Linux, macOS and Windows (Python 3.9 and 3.14). Eval suite 9/9 cases with claude plugin eval 2.1.281; with the skill against without it, the action and outcome cases score 1.00, 0.92, 1.00 and 1.00 against 0.29, 0.75, 0.00 and 0.67.

editwright.zip is the skill folder for claude.ai and Claude Desktop (Customize > Skills > Upload skill).