conorbronsdon/avoid-ai-writing

The final automated run for #295/#296 exposed a nonblocking diagnostic limitation separate from the unchanged-output blocker in #322. At source 0ae68f2fc3ddb166cbf8dc3156e2199cbd45bd13, the isolated Claude detect_only response calls unlocks a Tier 2 table entry and applies the cluster threshold to it. In references/patterns.md, unlock is a replacement for the listed Tier 2 entry unleash; it is not itself a word-table entry. Some responses also assign Moreover P1 while the skill's severity summary lists transition phrases under P2. The frozen detect scenario checks read-only scope, findings/assessment, and execution honesty. It does not validate every category or severity attribution, so those properties must not be inferred from its pass result. Follow-up: add model-only audit regressions that distinguish table entries from suggested replacements, preserve weak-match exceptions, and require category/severity claims to match the actual catalog. Keep this separate from deterministic d

Maintainer exception accepted, 2026-09-16. Conor explicitly approved landing #295 and #296 with the two recorded model-output failures: deleted source content and an omitted final artifact. The automated result remains 14/16, not a pass. These failures remain open follow-up work in #322; catalog-label variance remains in #323. Human evaluation is optional. This exception applies to this stack and does not waive future verification requirements. The earlier status below records the evaluation before this explicit acceptance. Follow-up status, 2026-09-16: The latest automated gate fails: 14 of 16 scenario responses met the frozen assertions. Claude deleted a source-internal sentence; MiMo omitted the final source in a technical-context no-op. These results remain blocking under the current gate; no human evaluation is required. Full follow-up evidence. The original false-edit report is repaired in the latest protected-content samples. The same frozen evaluation now exposes two remaining

wanna make sure if Chinese works out fine as well

Reserved for a first-time contributor. Please comment before starting and follow CONTRIBUTING.md's one-issue newcomer guidance. Prepared by Codex after reviewing #214; this is a separately scoped behavior fix. Problem and reproduction On main at 693438c019b959ec45f3b0d68fb3ebffb7f8680b, run: node -e 'console.log(require("./scripts/self-scan").scanFile("CHANGELOG.md"))' The result has chunked: 3, exemptIssues: 62, and topTypes: []. The file is currently under budget (34/40); the missing categories would make an over-budget diagnostic print none. scoreLongText() omits the category aggregation that the short-text path provides. Scope scripts/self-scan.js, a dedicated scripts/self-scan-diagnostics.test.js, package.json test wiring, and an Unreleased changelog entry. Aggregate category counts across all accepted chunks, then return the three most frequent types using the existing topTypes ordering. Keep scoring, chunk boundaries and budgets unchanged. Use a separate test file to avoid colli

$ node scripts/self-scan.js CHANGELOG.md 11617 60 35 40 Reproducing the exempt scan by hand isolates the score almost entirely to tier1 and tier2 hits from three release-note lines that list AI-tell vocabulary as unquoted prose (the exemption regex blanks only quoted spans, inline code, tables, and blockquotes): line 703: "39 new vocabulary entries ... bustling, intricate, complexities, ever-evolving, daunting, holistic, ..." line 716: "15 new word/phrase replacements: nuanced, crucial, multifaceted, ecosystem, myriad, plethora, ..." line 746: "5 new word table entries (nestled, vibrant, thriving, despite challenges, showcasing)" These are March 2026 entries (v1.1.0 to v2.0.0). Five points of headroom remain before npm run self-scan:check (part of detector-test.yml) fails on any future changelog entry that lists a few example words in prose, which is a normal shape for this project's release notes. Maintainer decision CONTRIBUTING.md says to leave publish