heznpc · heznpc · Jun 10, 2026 · Jun 10, 2026
@@ -434,3 +434,21 @@ By contributing, you agree that your contributions will be licensed under the [M
 Open a [Discussion](../../discussions) or file an issue. We're happy to help you get started!
 
 > 💡 **This document is in English only.** Want to translate it into your language? That PR is welcome too!
+
+## Native language reviewers
+
+The premium dictionaries are curated and LLM-audited, but **no locale has had
+a native-speaker pass yet** — and that's the highest-leverage contribution a
+non-coder can make here. One pass takes ~1–2 hours:
+
+1. Open `src/data/<your-locale>.json`. Every key is the English source string;
+   every value is the translation. Skim for wrong meaning, unnatural register,
+   brand terms that should stay English (Claude, Anthropic, MCP, Claude Code,
+   Managed Agents, …), and the same English term rendered inconsistently.
+2. Send corrections as a PR (JSON value edits only — CI guards the structure,
+   so you cannot break anything), or just open an issue listing them.
+3. Your locale's `_meta.nativeReview` flips to `"reviewed"` and you're credited
+   in the README's Terminology QA table.
+
+Claim a locale on [issue #202](https://github.com/heznpc/skillBridge/issues/202).
+The QA model around your review is described in `docs/TRANSLATION_QA.md`.
@@ -249,6 +249,44 @@ The full "things we will not do" list is kept public on purpose in [POSITIONING.
 
 > Want to add your language as Premium? Contribute a curated dictionary — see [CONTRIBUTING.md](CONTRIBUTING.md).
 
+### Terminology QA — how accuracy is enforced, not just promised
+
+New Academy content is covered by a standing pipeline, not by hand-checking:
+a CI watcher polls the live catalog twice a day and **fails loudly + opens an
+issue** the moment a course appears that the dictionaries don't cover; the
+course gets wired into all 11 premium dictionaries; structural CI gates
+(`check:i18n`, `check:dict-coverage`, `check:locales`) and a real-dictionary
+regression suite guard every merge after that. Proven turnaround: on
+**2026-06-10** the watcher flagged the brand-new *Claude Platform 101* course
+in the morning ([#196](https://github.com/heznpc/skillBridge/issues/196)) and
+all 11 locales were wired the same day
+([#201](https://github.com/heznpc/skillBridge/pull/201)).
+
+Beyond structure, dictionary *content* goes through layered review — CI gates
+catch shape/contamination drift on every PR, a full per-locale LLM audit runs
+before every store release (see `docs/TRANSLATION_QA.md`), and native-speaker
+review is the final layer:
+
+<!-- LOCALE_QA_START -->
+| Language | Code | Entries | Last curated | Last LLM audit | Native review |
+|---|---|---:|---|---|---|
+| 한국어 | `ko` | 1129 | 2026-04-02 | 2026-06-10 | 🙋 [recruiting](https://github.com/heznpc/skillBridge/issues/202) |
+| 日本語 | `ja` | 1129 | 2026-04-02 | 2026-06-10 | 🙋 [recruiting](https://github.com/heznpc/skillBridge/issues/202) |
+| 中文(简体) | `zh-CN` | 1129 | 2026-04-02 | 2026-06-10 | 🙋 [recruiting](https://github.com/heznpc/skillBridge/issues/202) |
+| 中文(繁體) | `zh-TW` | 1129 | 2026-04-02 | 2026-06-10 | 🙋 [recruiting](https://github.com/heznpc/skillBridge/issues/202) |
+| Español | `es` | 1129 | 2026-04-02 | 2026-06-10 | 🙋 [recruiting](https://github.com/heznpc/skillBridge/issues/202) |
+| Français | `fr` | 1129 | 2026-04-02 | 2026-06-10 | 🙋 [recruiting](https://github.com/heznpc/skillBridge/issues/202) |
+| Italiano | `it` | 1129 | 2026-06-03 | 2026-06-10 | 🙋 [recruiting](https://github.com/heznpc/skillBridge/issues/202) |
+| Deutsch | `de` | 1129 | 2026-04-02 | 2026-06-10 | 🙋 [recruiting](https://github.com/heznpc/skillBridge/issues/202) |
+| Português (BR) | `pt-BR` | 1129 | 2026-04-02 | 2026-06-10 | 🙋 [recruiting](https://github.com/heznpc/skillBridge/issues/202) |
+| Русский | `ru` | 1129 | 2026-04-02 | 2026-06-10 | 🙋 [recruiting](https://github.com/heznpc/skillBridge/issues/202) |
+| Tiếng Việt | `vi` | 1129 | 2026-04-02 | 2026-06-10 | 🙋 [recruiting](https://github.com/heznpc/skillBridge/issues/202) |
+<!-- LOCALE_QA_END -->
+
+🙋 **Native speakers wanted** — a first native pass on your locale takes
+~1–2 hours, needs no coding, and gets you credited here. See
+[#202](https://github.com/heznpc/skillBridge/issues/202).
+
 ## Privacy & Security
 
 SkillBridge is designed with privacy first:

@@ -0,0 +1,53 @@
+# Translation QA — the three assurance layers
+
+SkillBridge's core promise is AI-terminology fidelity. That promise is backed
+by three layers, each catching what the previous one structurally cannot.
+This doc is honest about what each layer does **and does not** guarantee.
+
+## Layer 1 — automated, every PR (CI)
+
+| Gate | Catches |
+|---|---|
+| `check:i18n` / `check:dict-coverage` | Key/section parity across all 11 dictionaries, `_meta.version` sync with the manifest |
+| `check:locales` | Cross-locale contamination — any locale sharing >8% of its long strings with another (the bug class behind the 2026-06 "Italian was 51% Spanish" incident) |
+| `check:glossary` / `validate` | `_protected` structure, value types, possibly-untranslated heuristics |
+| `tests/protected-terms.test.js` | Real-dictionary regression: ordinary prose in **all 11 locales** must survive protected-term restoration uncorrupted |
+| Academy drift watcher (cron, 2×/day) | A live course slug missing from `FLASHCARD_COURSE_MAP` — fails the run and auto-opens an issue |
+
+**Does not catch:** a translation that is structurally valid but semantically
+wrong ("uniquely wrong" values — e.g. `Slack → "Lento"` survived every
+structural gate until Layer 2 found it).
+
+## Layer 2 — full LLM audit, every store release
+
+Before each store submission, every premium dictionary gets a full-depth LLM
+review (one reviewer per locale over all ~1,100 entries, findings re-verified
+against the file before fixing — reviewer claims are *not* trusted blindly).
+The completed audit is stamped machine-readably in each dictionary:
+`_meta.lastAudited`. The README QA table is generated from that field by
+`npm run docs`, so the public table cannot drift from reality.
+
+The 2026-06-10 audit fixed 89 verified errors across 7 locales (brand-name
+mistranslations, a `responsskill` regex scar, "AI Fluency" rendered three
+different ways in one file, …) — see PRs #197/#199.
+
+**Does not catch:** native-register subtleties and domain idiom an LLM judges
+imperfectly. That's Layer 3.
+
+## Layer 3 — native-speaker review (recruiting)
+
+One native reviewer per locale does a first full pass (~1–2 hours, JSON value
+edits only; CI guards the structure so nothing can break). On completion the
+locale's `_meta.nativeReview` flips from `"recruiting"` to `"reviewed"` and the
+reviewer is credited in the README QA table.
+
+Status and sign-up: [issue #202](https://github.com/heznpc/skillBridge/issues/202).
+No locale has completed a native pass yet — the table says so honestly rather
+than implying review that hasn't happened.
+
+## Why there is no paid review / API gate
+
+SkillBridge is free forever for both users and the maintainer — no paid
+translation-QA API can sit in CI. The layered design above is the strongest
+assurance available under that constraint: structure by machine, semantics by
+release-time LLM audit, idiom by community native review.
@@ -11,6 +11,7 @@
  *   <!-- LANG_COUNT_START -->...<!-- LANG_COUNT_END -->
  *   <!-- PREMIUM_LANGS_START -->...<!-- PREMIUM_LANGS_END -->
  *   <!-- FEATURES_START -->...<!-- FEATURES_END -->
+ *   <!-- LOCALE_QA_START -->...<!-- LOCALE_QA_END -->  (per-locale QA table from src/data/*.json _meta)
  *
  * Usage:  node scripts/generate-docs.js
  */
@@ -126,6 +127,40 @@ function buildLangTagsHtml(langs) {
 // The script only updates version/lang counts/premium list; feature cards are maintained
 // manually but wrapped in markers so they *could* be generated in the future.
 
+// Per-locale QA table — generated from each dictionary's _meta so the README
+// table can never drift from the machine-readable QA state. `lastAudited` is
+// stamped by the pre-release LLM dictionary audit (see RELEASE_CHECKLIST step 0);
+// `nativeReview` flips to "reviewed" when a native speaker completes a pass
+// (recruiting: issue #202).
+function countLeaves(obj) {
+  let n = 0;
+  for (const [k, v] of Object.entries(obj)) {
+    if (k === '_meta' || k === '_protected') continue;
+    if (typeof v === 'string') n++;
+    else if (v && typeof v === 'object' && !Array.isArray(v)) n += countLeaves(v);
+  }
+  return n;
+}
+function buildLocaleQaTable() {
+  const rows = [];
+  for (const { code, label } of premiumLangs) {
+    const dict = JSON.parse(fs.readFileSync(path.join(ROOT, `src/data/${code}.json`), 'utf8'));
+    const m = dict._meta || {};
+    const review =
+      m.nativeReview === 'reviewed'
+        ? '✅ reviewed'
+        : '🙋 [recruiting](https://github.com/heznpc/skillBridge/issues/202)';
+    rows.push(
+      `| ${label} | \`${code}\` | ${countLeaves(dict)} | ${m.lastUpdated || '—'} | ${m.lastAudited || '—'} | ${review} |`,
+    );
+  }
+  return [
+    '| Language | Code | Entries | Last curated | Last LLM audit | Native review |',
+    '|---|---|---:|---|---|---|',
+    ...rows,
+  ].join('\n');
+}
+
 // ---------------------------------------------------------------------------
 // 4. Replace between markers
 // ---------------------------------------------------------------------------
@@ -156,6 +191,7 @@ let readme = fs.readFileSync(readmePath, 'utf8');
 
 readme = replaceBetweenMarkers(readme, 'VERSION', `v${version}`);
 readme = replaceBetweenMarkers(readme, 'LANG_COUNT', `${translatedCount} languages`);
+readme = replaceBetweenMarkers(readme, 'LOCALE_QA', '\n' + buildLocaleQaTable() + '\n');
 
 fs.writeFileSync(readmePath, readme, 'utf8');
 

@@ -3,7 +3,9 @@
     "lang": "de",
     "langName": "Deutsch",
     "version": "3.5.40",
-    "lastUpdated": "2026-04-02"
+    "lastUpdated": "2026-04-02",
+    "lastAudited": "2026-06-10",
+    "nativeReview": "recruiting"
   },
   "ui": {
     "Anthropic Academy": "Anthropic Academy",

@@ -3,7 +3,9 @@
     "lang": "es",
     "langName": "Español",
     "version": "3.5.40",
-    "lastUpdated": "2026-04-02"
+    "lastUpdated": "2026-04-02",
+    "lastAudited": "2026-06-10",
+    "nativeReview": "recruiting"
   },
   "ui": {
     "Anthropic Academy": "Anthropic Academy",

@@ -3,7 +3,9 @@
     "lang": "fr",
     "langName": "Français",
     "version": "3.5.40",
-    "lastUpdated": "2026-04-02"
+    "lastUpdated": "2026-04-02",
+    "lastAudited": "2026-06-10",
+    "nativeReview": "recruiting"
   },
   "ui": {
     "Anthropic Academy": "Anthropic Academy",

@@ -4,7 +4,9 @@
     "langName": "Italiano",
     "version": "3.5.40",
     "lastUpdated": "2026-06-03",
-    "translation_provenance": "Re-translated from the English source via Google Translate (2026-06), with brand/technical terms restored to canonical English; Italian↔Spanish overlap is now 0.1% (parity with the other 10 locales). Earlier v1 was a Spanish-derived regex transform — fully replaced. Structural keys must remain identical to other premium dictionaries (enforced by scripts/check-dict-coverage.js); native Italian PRs welcome."
+    "translation_provenance": "Re-translated from the English source via Google Translate (2026-06), with brand/technical terms restored to canonical English; Italian↔Spanish overlap is now 0.1% (parity with the other 10 locales). Earlier v1 was a Spanish-derived regex transform — fully replaced. Structural keys must remain identical to other premium dictionaries (enforced by scripts/check-dict-coverage.js); native Italian PRs welcome.",
+    "lastAudited": "2026-06-10",
+    "nativeReview": "recruiting"
   },
   "ui": {
     "Anthropic Academy": "Anthropic Academy",

@@ -3,7 +3,9 @@
     "lang": "ja",
     "langName": "日本語",
     "version": "3.5.40",
-    "lastUpdated": "2026-04-02"
+    "lastUpdated": "2026-04-02",
+    "lastAudited": "2026-06-10",
+    "nativeReview": "recruiting"
   },
   "ui": {
     "Anthropic Academy": "Anthropic Academy",

@@ -3,7 +3,9 @@
     "lang": "ko",
     "langName": "한국어",
     "version": "3.5.40",
-    "lastUpdated": "2026-04-02"
+    "lastUpdated": "2026-04-02",
+    "lastAudited": "2026-06-10",
+    "nativeReview": "recruiting"
   },
   "ui": {
     "Anthropic Academy": "Anthropic 아카데미",

@@ -3,7 +3,9 @@
     "lang": "pt-BR",
     "langName": "Português (BR)",
     "version": "3.5.40",
-    "lastUpdated": "2026-04-02"
+    "lastUpdated": "2026-04-02",
+    "lastAudited": "2026-06-10",
+    "nativeReview": "recruiting"
   },
   "ui": {
     "Anthropic Academy": "Anthropic Academy",

@@ -3,7 +3,9 @@
     "lang": "ru",
     "langName": "Русский",
     "version": "3.5.40",
-    "lastUpdated": "2026-04-02"
+    "lastUpdated": "2026-04-02",
+    "lastAudited": "2026-06-10",
+    "nativeReview": "recruiting"
   },
   "ui": {
     "Anthropic Academy": "Anthropic Academy",

@@ -3,7 +3,9 @@
     "lang": "vi",
     "langName": "Tiếng Việt",
     "version": "3.5.40",
-    "lastUpdated": "2026-04-02"
+    "lastUpdated": "2026-04-02",
+    "lastAudited": "2026-06-10",
+    "nativeReview": "recruiting"
   },
   "ui": {
     "Anthropic Academy": "Anthropic Academy",

@@ -3,7 +3,9 @@
     "lang": "zh-CN",
     "langName": "简体中文",
     "version": "3.5.40",
-    "lastUpdated": "2026-04-02"
+    "lastUpdated": "2026-04-02",
+    "lastAudited": "2026-06-10",
+    "nativeReview": "recruiting"
   },
   "ui": {
     "Anthropic Academy": "Anthropic Academy",

@@ -3,7 +3,9 @@
     "lang": "zh-TW",
     "langName": "中文(繁體)",
     "version": "3.5.40",
-    "lastUpdated": "2026-04-02"
+    "lastUpdated": "2026-04-02",
+    "lastAudited": "2026-06-10",
+    "nativeReview": "recruiting"
   },
   "ui": {
     "Anthropic Academy": "Anthropic Academy",

@@ -41,6 +41,15 @@ Everything code-side is ready and pre-built.
 
 ## What needs your hands
 
+### 0. Pre-release dictionary audit (LLM) — release convention
+
+Before every store submission, run the full per-locale LLM dictionary audit
+(one reviewer per premium locale over all entries; re-verify every finding
+against the file before fixing — the 2026-06-10 audit caught `Slack → "Lento"`
+-class errors that all structural gates miss). After fixes land, stamp each
+dictionary's `_meta.lastAudited` and run `npm run docs` so the README QA table
+reflects it. Three-layer QA model: `docs/TRANSLATION_QA.md`.
+
 ### 1. Icon — resolved
 
 Status: **resolved**. The non-infringing icon shipped in v3.5.35 (a rising