diff --git a/CHANGELOG.md b/CHANGELOG.md
index 5fbfc7a..e55b813 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -5,6 +5,10 @@ Conventional Commits by release-please.
 
 ## Unreleased
 
+- Fixing a word's language no longer makes the word disappear when the model answers with the
+  same word in the same language ("Already in your list"): the word stays in your list. Before,
+  it went to Recently deleted.
+
 - Privacy policy version 6: "What websites can see" now names every way a page can hide
   text that Kotiko still swaps (a transparent text colour, text turned away), and says a
   disguised word's card can open on a click as well as a resting pointer.
diff --git a/README.md b/README.md
index 83c5821..ce99eba 100644
--- a/README.md
+++ b/README.md
@@ -85,7 +85,7 @@ data folders and the service file.
 ## Privacy
 
 The text of the pages you read never leaves your browser. A word you add goes only to the AI
-service you chose, or to your own server. Kotiko has no servers of its own and collects
+service you chose, or to your own server, and, for its pronunciation, to English Wiktionary. Kotiko has no servers of its own and collects
 nothing. The [privacy policy](https://kotiko.org/privacy/) (its source is
 [docs/privacy/en.md](docs/privacy/en.md)) says exactly what is kept and sent, and
 [docs/privacy/inventory.md](docs/privacy/inventory.md) lists every request with the code
diff --git a/docs/security/review-v1.0.0.md b/docs/security/review-v1.0.0.md
index 53cebf7..56b88a7 100644
--- a/docs/security/review-v1.0.0.md
+++ b/docs/security/review-v1.0.0.md
@@ -134,8 +134,11 @@ rc.4 zips rebuild byte for byte from the tag.
 with issues: SCR-836 (treat a transparent text colour and text turned away as hidden) and
 SCR-837 (read the report the way a CommonMark renderer shows it). The privacy policy is
 now version 6, which names every way a page can hide text that Kotiko still swaps.
-v1.0.0 differs from rc.4 only in documentation (this report, the privacy policy's wording
-and other docs) and the release notes.
+v1.0.0 differs from rc.4 in documentation (this report, the privacy policy's wording and
+other docs), the release notes, and one fix from the launch checks that touches no security
+control: fixing a word's language no longer deletes the word when the new add lands on it
+(`retireReplaced` in `extension/background.js` keeps that word; launch check F1, with a
+regression test).
 
 One slip: a misquoted test setting sent two of F's test lookups ("shukran", "gato") to
 OpenRouter's public endpoint, with no key, token or identity. OpenRouter refused them
diff --git a/extension/background.js b/extension/background.js
index f6fd3b8..60ed3db 100644
--- a/extension/background.js
+++ b/extension/background.js
@@ -1995,7 +1995,10 @@ async function relang(m) {
 
 async function retireReplaced(job) {
   const { id, key } = job.replaces;
-  const res = await undoWord(id, key);
+  // The new add can land on the very word the old one made ("Already in your list", when the
+  // model answers the same word in the same language): that word stays.
+  const keep = new Set((job.results ?? []).map((r) => r.wordId).filter(Boolean));
+  const res = await undoWord(id, key, { keep });
   if (res.ok) await queue.patch(id, { replacedBy: job.id });
 }
 
@@ -2065,8 +2068,8 @@ const jobWords = async (id, key) => {
 };
 const writeOps = async (ops) => (await wordRoutes["words.write"].run({ ops })).results;
 
-async function undoWord(id, key) {
-  const records = (await jobWords(id, key)).filter((r) => r.result !== "unchanged" && r.undo !== "done");
+async function undoWord(id, key, { keep = new Set() } = {}) {
+  const records = (await jobWords(id, key)).filter((r) => r.result !== "unchanged" && r.undo !== "done" && !keep.has(r.wordId));
   if (!records.length) return { ok: true };
   await queue.setUndo(id, key, { undo: "pending", undoError: null });
   const ops = records.map((r) =>
diff --git a/site/src/content/docs/index.mdx b/site/src/content/docs/index.mdx
index e2d2974..76c1ce7 100644
--- a/site/src/content/docs/index.mdx
+++ b/site/src/content/docs/index.mdx
@@ -48,7 +48,7 @@ Kotiko never assumes you read English.
 
 <CardGrid>
   <Card title="Type a word and its meaning" icon="pencil">
-    Write `perro = dog` and it's yours. No account, no key, nothing leaves your browser.
+    Write `perro = dog` and it's yours. No account, no key, no AI service.
   </Card>
   <Card title="Connect a free AI key" icon="rocket">
     With a free OpenRouter key, ask in your own words: "how do you say cat in Japanese".
@@ -71,7 +71,8 @@ Until the store listings are live, you can [install Kotiko from its source](/ins
 ## Private by design
 
 - The text of the pages you read never leaves your browser.
-- A word you add goes only to the AI service you chose, or to your own server.
+- A word you add goes only to the AI service you chose, or to your own server, and, for its
+  pronunciation, to English Wiktionary.
 - Kotiko has no servers of its own, no accounts, no ads and no tracking. Neither does this site.
 
 Read the [privacy policy](/privacy/): it takes about two minutes.
diff --git a/site/src/content/docs/start.md b/site/src/content/docs/start.md
index 02fbc8d..0b24ef8 100644
--- a/site/src/content/docs/start.md
+++ b/site/src/content/docs/start.md
@@ -23,7 +23,9 @@ Pick one of three ways. You can change it at any time in **Settings**, **Word lo
 ### No key: type words with their meaning
 
 Select **Skip: I'll type meanings myself**. You add words as `word = meaning`, like
-`perro = dog` or `спасибо (spasibo) = thanks`. Nothing leaves your browser. This always works,
+`perro = dog` or `спасибо (spasibo) = thanks`. No AI service is asked. For a language with
+word stress, such as Russian or Spanish, Kotiko looks up only the word on English
+Wiktionary to get its pronunciation right ([privacy policy](/privacy/)). This always works,
 even with a key set up.
 
 ### A free OpenRouter key (recommended)
diff --git a/test/bg/local-mode.test.mjs b/test/bg/local-mode.test.mjs
index ae9f4e6..2f43933 100644
--- a/test/bg/local-mode.test.mjs
+++ b/test/bg/local-mode.test.mjs
@@ -128,6 +128,22 @@ describe("a fresh install keeps words in this browser (slice 11)", () => {
     assert.equal(bg.store.addJobs.find((j) => j.id === first.job.id).replacedBy, null);
   });
 
+  // Launch check F1: the model can answer the same word in the same language ("Already in
+  // your list"); retiring the old add must not delete the word the new one landed on.
+  test("another language that comes back as the same word keeps the word", async () => {
+    const bg = loadBackground({ local: { baseLangs: ["en"] } });
+    await bg.k.ready();
+    const first = await bg.send({ type: "add", text: "gato = cat", hintLang: "es" }, POPUP);
+    await bg.until(() => bg.store.addJobs?.find((j) => j.id === first.job.id && j.state === "done"));
+    const relang = await bg.send({ type: "jobs.relang", id: first.job.id, key: "es\u001fgato", lang: "es" }, POPUP);
+    assert.equal(relang.ok, true);
+    const next = await bg.until(() => bg.store.addJobs?.find((j) => j.id === relang.newId && j.state === "done"));
+    assert.equal(next.results[0].result, "unchanged", "the new add found the same word");
+    await bg.until(() => bg.store.addJobs?.find((j) => j.id === first.job.id && j.replacedBy === relang.newId));
+    const active = (await (await bg.k.getStore()).list()).map((w) => `${w.lang}:${w.native}`);
+    assert.deepEqual(active, ["es:gato"], "the word is still in the list");
+  });
+
   test("Focus on one language decides a 'native = meaning' word's language", async () => {
     const bg = loadBackground({ local: { baseLangs: ["en"] } });
     await bg.k.ready();