diff --git a/.changeset/settings-getmany-all-or-nothing-doc.md b/.changeset/settings-getmany-all-or-nothing-doc.md
new file mode 100644
index 0000000000..8638452552
--- /dev/null
+++ b/.changeset/settings-getmany-all-or-nothing-doc.md
@@ -0,0 +1,36 @@
+---
+"@objectstack/service-settings": patch
+---
+
+docs(service-settings): state `getMany`'s all-or-nothing key validation on the declaration that owns it (#11680)
+
+Documentation and a pin. **No behaviour change** — the accept set and every
+resolved value are byte-identical.
+
+`SettingsService.getMany` validates **every** requested key against the
+namespace manifest before it reads a single env override and before it loads a
+single row, so one undeclared key rejects the whole call with
+`UnknownKeyError` (`code: 'SETTINGS_UNKNOWN_KEY'`) and the caller receives
+nothing — not the subset it was entitled to. N per-key `get()` calls behave
+differently on exactly that input: each declared key still answers, and only
+the undeclared one throws.
+
+The doc comment was otherwise detailed — it explained the grouped row load and
+the env-override ordering, and claimed row-for-row equivalence with per-key
+`get` "BY CONSTRUCTION" — but never drew this line. That equivalence claim
+holds for every key that *resolves* and not for the refusal, so a batched
+consumer had to rediscover the rule from a test. `resolveLocalizationContext`
+was the first to inherit it and had to record the consequence locally: a host
+registering a **partial** `localization` manifest loses all its keys at once
+and drops to a shorter cascade, where the per-key path would still have
+resolved the declared ones.
+
+`getMany`'s doc comment now states the rule, its blast radius, why validating
+ahead of the grouped walk makes the refusal independent of key order and scope
+grouping, and what a caller on a partial manifest should expect.
+
+The pre-existing pin asserted only `rejects.toThrow(/nope/)` — green whatever
+the blast radius is. A sibling pin now asserts the property instead: the error
+envelope (`code: 'SETTINGS_UNKNOWN_KEY'`), that **zero** rows were loaded
+(the refusal is up-front, with the undeclared key last in the request), and
+the contrast that per-key `get()` still answers each declared key.
diff --git a/packages/services/service-settings/src/settings-getmany.test.ts b/packages/services/service-settings/src/settings-getmany.test.ts
index b8e2d321a1..db0cee8f32 100644
--- a/packages/services/service-settings/src/settings-getmany.test.ts
+++ b/packages/services/service-settings/src/settings-getmany.test.ts
@@ -20,6 +20,7 @@
 
 import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
 import { SettingsService } from './settings-service.js';
+import { UnknownKeyError } from './settings-service.types.js';
 
 // WHERE-matcher gate: implement exactly the combinators the service emits and
 // THROW on the rest — a bare field-equality read of `$or` would silently match
@@ -118,6 +119,42 @@ describe('[#10826] SettingsService.getMany', () => {
     await expect(svc.get('localization', 'nope')).rejects.toThrow(/nope/);
   });
 
+  // [#11680] The rule the doc comment now states, pinned as a PROPERTY rather
+  // than as "it throws": the refusal is TOTAL (no partial Record), it lands
+  // BEFORE any row load, and it is the one input on which `getMany` and N
+  // per-key `get()` calls part ways. The sibling above asserts only that the
+  // message names the bad key — which stays green whatever the blast radius is.
+  it('one undeclared key rejects the WHOLE call — before any row load, no partial result', async () => {
+    const { svc, engine } = await makeService();
+    engine.find.mockClear();
+
+    const err: unknown = await svc
+      .getMany('localization', ['timezone', 'currency', 'nope'])
+      .then(() => null, (e: unknown) => e);
+
+    // The envelope, not just the throw. This is a service-layer error class:
+    // it carries `code` and no `status` (no HTTP boundary here), so `code` is
+    // the whole machine-readable envelope there is to assert.
+    expect(err).toBeInstanceOf(UnknownKeyError);
+    expect((err as UnknownKeyError).code).toBe('SETTINGS_UNKNOWN_KEY');
+    expect((err as Error).message).toMatch(
+      /Key 'nope' is not declared in manifest 'localization'/,
+    );
+
+    // TOTAL and UP-FRONT: the two DECLARED keys were neither answered nor even
+    // loaded — validation runs ahead of the grouped `loadRows`.
+    expect(engine.find).toHaveBeenCalledTimes(0);
+
+    // ...and here is the non-equivalence: per-key `get()` still answers every
+    // declared key on the same input; only the undeclared one throws. Asserted
+    // on the resolved cascade LAYER, not on the literal — this fixture stores
+    // JSON text in `value` while the service persists values verbatim, so a
+    // literal here would pin the fixture's encoding rather than the rule.
+    expect((await svc.get('localization', 'timezone')).source).toBe('global');
+    expect((await svc.get('localization', 'currency')).source).toBe('tenant');
+    await expect(svc.get('localization', 'nope')).rejects.toBeInstanceOf(UnknownKeyError);
+  });
+
   it('getNamespace resolves through the grouped path with unchanged answers', async () => {
     const { svc, engine } = await makeService();
     const ctx = { userId: 'u1' };
diff --git a/packages/services/service-settings/src/settings-service.ts b/packages/services/service-settings/src/settings-service.ts
index cf4a18920d..f28fa8f381 100644
--- a/packages/services/service-settings/src/settings-service.ts
+++ b/packages/services/service-settings/src/settings-service.ts
@@ -1218,6 +1218,40 @@ export class SettingsService {
    * the env-override branch, the scope→userId mapping, and the cascade are
    * the same code ({@link resolveKeyFromRows} is extracted from `get`, not
    * copied). Nothing is cached; nothing survives the call.
+   *
+   * ## Key validation is UP-FRONT and TOTAL
+   *
+   * That equivalence covers every key that RESOLVES. It does not cover the
+   * refusal, and this is the one respect in which `getMany` is not N `get`
+   * calls — so it is stated here rather than left to be rediscovered from a
+   * test.
+   *
+   * EVERY requested key is checked against the namespace's manifest before a
+   * single env override is read and before any row is loaded. One undeclared
+   * key therefore rejects the WHOLE call — {@link UnknownKeyError}, `code:
+   * 'SETTINGS_UNKNOWN_KEY'` — and the caller receives NOTHING: no partial
+   * `Record`, not even the subset it was entitled to. N per-key {@link get}
+   * calls part ways on exactly this input: each declared key still answers,
+   * and only the undeclared one throws. Same error class, same code; the
+   * blast radius is what differs. (An unregistered NAMESPACE is refused first
+   * and identically to `get`: {@link UnknownNamespaceError}.)
+   *
+   * Validating before the grouped walk rather than inside it is what makes
+   * the refusal independent of key order, of scope grouping, and of which
+   * keys happened to carry an env override — the call either refuses or
+   * answers all of them, never something in between. `setMany` pre-flights
+   * its whole patch the same way.
+   *
+   * What it costs a caller: against a host that registers a PARTIAL manifest,
+   * a batched consumer loses ALL of its keys at once and must degrade for the
+   * whole set — it cannot fall back key by key. `resolveLocalizationContext`
+   * is the first consumer to inherit this and records its own degradation
+   * locally: a partial `localization` manifest drops it to a shorter cascade
+   * (no `global` scope layer, no `OS_LOCALIZATION_*` override) where the
+   * per-key path would still have resolved the declared keys. A caller that
+   * cannot afford the rule should intersect `keys` with the manifest itself,
+   * or read per key. {@link getNamespace} can never trip it — it passes
+   * exactly the registered keys.
    */
   async getMany(
     namespace: string,