diff --git a/.changeset/8181-permission-door-read-decoration-pin.md b/.changeset/8181-permission-door-read-decoration-pin.md
new file mode 100644
index 0000000000..a9ca6b340d
--- /dev/null
+++ b/.changeset/8181-permission-door-read-decoration-pin.md
@@ -0,0 +1,8 @@
+---
+---
+
+Pin the permission door's package save against the framework's read decorations
+(objectui#8181). The inline `stripReadDecorations` at
+`PermissionMatrixEditor`'s draft unwrap was the one strip in the sweep that no
+test asserted — measured by removing it, which turned nothing red. Test only;
+no package is released by this change.
diff --git a/packages/app-shell/src/views/metadata-admin/PermissionMatrixEditor.readDecorationStrip.test.tsx b/packages/app-shell/src/views/metadata-admin/PermissionMatrixEditor.readDecorationStrip.test.tsx
new file mode 100644
index 0000000000..ec81acb0cf
--- /dev/null
+++ b/packages/app-shell/src/views/metadata-admin/PermissionMatrixEditor.readDecorationStrip.test.tsx
@@ -0,0 +1,179 @@
+// Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+
+/**
+ * The permission door's package save carries no read decorations
+ * (objectui#8181).
+ *
+ * ## Why this file exists after the strip already landed
+ *
+ * objectui#8181 asked for a read-decoration strip at three consumers. All
+ * three have one on `main`. Two of them — `ObjectHooksPanel` and
+ * `StudioDesignSurface` — delegate to the hoisted `extractDraftBody` in
+ * `@object-ui/data-objectstack`, so deleting their strip is caught by
+ * `draft-envelope.test.ts`. THIS consumer is the exception: it does not call
+ * the hoisted reader, it calls `stripReadDecorations` inline at its own unwrap.
+ *
+ * Measured before this file was written: removing that inline strip — leaving
+ * the unwrap, i.e. restoring the exact `((pendingDraft as any).item ??
+ * pendingDraft)` shape the card names — turned NOTHING red across
+ * `PermissionMatrixEditor.*` and both `data-objectstack` draft suites. A lit
+ * control run at the same line (dropping the pending draft entirely) DID turn
+ * one test red, so the path is exercised; it was the decorations specifically
+ * that nothing asserted. That is what this file closes: the strip was present
+ * and unguarded, one edit away from silently coming back.
+ *
+ * ## Why the `.catch` arm and not the everyday one
+ *
+ * ⚠️ The success arm cannot carry decorations and is not evidence either way.
+ * `doSave` re-reads `layered` and merges through `mergePermissionSlice`, which
+ * starts from `{ ...base }` and copies only `EDITOR_AUTHORED_KEYS` out of the
+ * edited draft — so when the fresh read succeeds, `base` is the RAW `effective`
+ * layer and every decoration is dropped by the merge whether or not the unwrap
+ * stripped. A pin written there would pass with the strip deleted.
+ *
+ * The leak the card names is the other arm: `client.layered(...).catch(() =>
+ * null)` falls back to `base = payload`, the draft body itself, and
+ * `mergePermissionSlice` then spreads it wholesale into `client.save`. The card
+ * called this out as "a failure-path-only leak and it has not been exercised".
+ * It is exercised here.
+ *
+ * ## What the far side does, stated so nobody re-derives it
+ *
+ * The server does not 400 on this today: `saveMetaItem` strips read decorations
+ * on ingress, ahead of its schema gate. That is a mitigation in the framework,
+ * not a licence for this client to PUT a body its own spec rejects —
+ * `PermissionSetSchema.safeParse(body + _diagnostics)` answers
+ * `unrecognized_keys` at the root. AGENTS.md #0.1: fix it at the producer.
+ */
+
+import '@testing-library/jest-dom/vitest';
+import { afterEach, describe, expect, it, vi } from 'vitest';
+import { cleanup, fireEvent, render, screen, waitFor, within } from '@testing-library/react';
+import { MemoryRouter } from 'react-router-dom';
+
+/** Exactly what the framework attaches to a `?state=draft` read. */
+const DECORATIONS = { _diagnostics: { valid: true, errors: [] }, _draft: true };
+/**
+ * ADR-0010 carriers. They share the underscore spelling, the closed schemas
+ * DECLARE them, and `METADATA_READ_DECORATIONS` deliberately omits them — so
+ * they must ride through untouched. Asserted here as a live control: they are
+ * the proof that an underscore-prefixed key from the draft body really does
+ * reach the save on this arm, which is what makes the absence of
+ * `_diagnostics` a strip rather than a blanket underscore filter.
+ */
+const PROTECTION = { _provenance: 'package', _packageId: 'app.a' };
+
+/** Authored in a previous package save, staged in the draft, never published. */
+const AUTHORED_POLICY = {
+  name: 'qa_st2_rls',
+  object: 'a_account',
+  operation: 'all',
+  using: 'owner_id == current_user.id',
+  enabled: true,
+};
+
+const PUBLISHED = {
+  name: 'sales_perms',
+  label: 'Sales',
+  objects: {
+    a_account: { allowRead: true, allowCreate: true },
+    a_contact: { allowRead: true },
+  },
+  fields: {},
+};
+
+interface Server {
+  saved: Array<Record<string, unknown>>;
+  /** `layered` calls seen so far — #1 is the load, #2 is `doSave`'s re-read. */
+  layeredCalls: number;
+}
+
+/**
+ * A client whose `layered` succeeds for the LOAD and then fails for `doSave`'s
+ * re-read — the arm the card names. Failing both would leave the editor with no
+ * published baseline at all and test a different screen.
+ */
+function makeClient(server: Server) {
+  return {
+    layered: async () => {
+      server.layeredCalls += 1;
+      if (server.layeredCalls >= 2) throw new Error('layered read failed');
+      return { effective: PUBLISHED, code: null, overlay: null, overlayScope: null };
+    },
+    // The decorated envelope, exactly as `decorateMetadataItem` serves it.
+    getDraft: async () => ({
+      type: 'permission',
+      name: 'sales_perms',
+      item: {
+        ...PUBLISHED,
+        rowLevelSecurity: [AUTHORED_POLICY],
+        ...PROTECTION,
+        ...DECORATIONS,
+      },
+    }),
+    list: async (type: string) =>
+      type === 'object' ? [{ item: { name: 'a_account' } }, { item: { name: 'a_contact' } }] : [],
+    get: async () => null,
+    save: async (_type: string, _name: string, payload: Record<string, unknown>) => {
+      server.saved.push(payload);
+      return payload;
+    },
+  } as any;
+}
+
+let clientImpl: any;
+
+vi.mock('./useMetadata', () => ({
+  useMetadataClient: () => clientImpl,
+  useMetadataTypes: () => ({
+    loading: false,
+    error: null,
+    entries: [{ type: 'permission', label: 'Permission', allowOrgOverride: true }],
+  }),
+}));
+
+vi.mock('./AssignedUsersSection', () => ({ AssignedUsersSection: () => null }));
+
+import { PermissionMatrixEditPage } from './PermissionMatrixEditor';
+
+afterEach(cleanup);
+
+describe('PermissionMatrixEditPage — read decorations never reach the save (objectui#8181)', () => {
+  it('drops `_diagnostics` / `_draft` even when the save-time layered re-read fails', async () => {
+    const server: Server = { saved: [], layeredCalls: 0 };
+    clientImpl = makeClient(server);
+
+    render(
+      <MemoryRouter>
+        <PermissionMatrixEditPage type="permission" name="sales_perms" packageId="app.a" />
+      </MemoryRouter>,
+    );
+    await screen.findByText('a_account');
+
+    // An ordinary edit — the author is not touching RLS or anything internal.
+    const row = screen.getByText('a_account').closest('tr')!;
+    fireEvent.click(within(row).getByRole('button', { name: 'None' }));
+
+    fireEvent.click(screen.getByRole('button', { name: /^Save$/ }));
+    await waitFor(() => expect(server.saved).toHaveLength(1));
+    const body = server.saved[0] as Record<string, any>;
+
+    // ── CONTROLS. Every absence below is meaningless without these. ──
+
+    // We are on the `.catch` arm: the re-read was attempted and it failed.
+    expect(server.layeredCalls).toBe(2);
+    // The DRAFT BODY is the merge base, not the published record — proved by a
+    // key only the draft carries. Without this the body under test could be the
+    // published baseline, which never had decorations to lose.
+    expect(body.rowLevelSecurity).toEqual([AUTHORED_POLICY]);
+    // The author's edit really is in the body being PUT.
+    expect(body.objects.a_account).toEqual({});
+    // Underscore-prefixed keys from the draft DO reach the save on this arm…
+    expect(body._provenance).toBe('package');
+    expect(body._packageId).toBe('app.a');
+
+    // ── THE PIN. Restoring the pre-fix unwrap turns exactly these red. ──
+    expect(body).not.toHaveProperty('_diagnostics');
+    expect(body).not.toHaveProperty('_draft');
+  });
+});