buvard 0.0.0-stage → 0.1.0

package/AGENTS.md

@@ -0,0 +1,41 @@
+# buvard
+
+**L'anti-slop qui tache.** Blotting paper for AI-flavored prose: a weighted,
+multi-score style linter for French and English slop patterns. GPL-3.0,
+repo `git.nech.pl/pln/buvard` (visibility toggle is the owner's call).
+
+## Where things stand (2026-11-07)
+
+- **Catalog v2** in `lib/catalog.mjs` — weighted patterns clustered from five
+  anti-slop sources, provenance tag per pattern. Barème: 100 base, lexique -5,
+  structure -8, servile/meta -10; punctuation budgets (1 tiret cadratin/500w,
+  1 !/1000w). **Voice rule: hits are review candidates, never verdicts — the
+  author's voice always wins.** buvard is a style linter, never an AI detector.
+- **`bin/buvard.js`** — CLI (`check`, `--json`), frontmatter stripped. Tests:
+  `npm test` (node:test, 6/6 passing).
+- **`docs/research.md`** — the clustering study. **`docs/research-litreview.md`**
+  — literature review (2026-11-07): detection literature unreliable on short/
+  paraphrased text; 61.22% FP rate of commercial detectors on non-native essays
+  (arxiv 2304.02819) is the core argument for staying a linter; French is open
+  territory (MULTITuDE omits it); ends with **10 falsifiable recommendations
+  for catalog v3** — read it before touching the catalog.
+- Extracted from the www multi-score suite; www still runs its own
+  `scripts/slop.mjs` — swap to the buvard dependency once on npm.
+
+## Pending (owner)
+
+1. Flip GitLab visibility to public (same toggle as devname).
+2. `npm login` then `npm publish` (name free on npm + PyPI; PyPI registration
+   not done yet).
+3. Catalog v3 from the lit review's recommendations — semver with per-rule
+   provenance (source URL + date + severity), Gutenberg pre-2022 FP corpus with
+   a <2% per-rule FP gate, language-specific dash budgets (EN 1/500w,
+   FR 1/1000w), "X, so Y" stays at folklore weight until a second source.
+
+## Conventions
+
+- GPL-3.0-only; no AI fingerprints in commits; tiny atomic commits.
+- Node >= 18, ESM (`"type": "module"`), zero runtime dependencies.
+- Regex catalog is deliberately high-precision/low-recall: règle de trois,
+  synonyme acrobatique, sentence-length monotony stay manual checks (see
+  SKILL.md's known-gaps list).