diff --git a/locators/models_locators.py b/locators/models_locators.py
index 91fd421..e0ece4e 100644
--- a/locators/models_locators.py
+++ b/locators/models_locators.py
@@ -50,6 +50,11 @@ class ModelsLocators:
     START_EVALUATION_LINK = "text=Start Evaluation"
     # Platform UI uses 'Invite Evaluators' (was 'Auditors' in earlier copy).
     INVITE_AUDITORS_LINK = "text=/Invite (Auditors|Evaluators)/i"
+    # The invitation dialog AuditorInvitation.tsx opens per-version. Its
+    # Dialog.Content title is 'Invite Evaluator' or 'Invite Evaluator for
+    # Version <n>' — match the stable prefix.
+    # Added for test-sync/ParakhAI-frontend-pr443 (a11y label fix).
+    INVITE_AUDITOR_DIALOG = "[role='dialog']:has-text('Invite Evaluator')"
     # Playwright's `text=` engine treats commas as part of the literal string,
     # so multi-selector fallbacks must use CSS-only forms (`:has-text()`).
     DATE_UPDATED_COL = "text=DATE UPDATED"
diff --git a/pages/models_page.py b/pages/models_page.py
index 8ebebb2..5d4ec62 100644
--- a/pages/models_page.py
+++ b/pages/models_page.py
@@ -135,6 +135,31 @@ def is_start_evaluation_visible(self) -> bool:
     def is_invite_auditors_visible(self) -> bool:
         return self.is_visible(self.INVITE_AUDITORS_LINK)
 
+    def click_invite_auditors(self) -> bool:
+        """Click the first 'Invite Evaluators' link/button (per model version)
+        and wait for the AuditorInvitation dialog to open.
+
+        Returns True if the dialog opened, False if there was nothing to
+        click (e.g. no versions on this model) — callers should skip rather
+        than fail on False, since that's a data gap, not a product bug.
+        """
+        link = self.page.locator(self.INVITE_AUDITORS_LINK).first
+        try:
+            link.wait_for(state="visible", timeout=8_000)
+        except Exception:
+            return False
+        link.click()
+        try:
+            self.page.locator(ModelsLocators.INVITE_AUDITOR_DIALOG).first.wait_for(
+                state="visible", timeout=self.timeout
+            )
+        except Exception:
+            return False
+        return True
+
+    def is_invite_auditor_dialog_visible(self) -> bool:
+        return self.is_visible(ModelsLocators.INVITE_AUDITOR_DIALOG)
+
     def is_past_evaluations_visible(self) -> bool:
         return self.is_visible(self.PAST_EVALUATIONS_HEADING)
 
diff --git a/tests/accessibility/test_auditor_invitation_a11y.py b/tests/accessibility/test_auditor_invitation_a11y.py
new file mode 100644
index 0000000..becaf82
--- /dev/null
+++ b/tests/accessibility/test_auditor_invitation_a11y.py
@@ -0,0 +1,101 @@
+"""
+Pending coverage for CivicDataLab/ParakhAI-frontend#443
+("fix(a11y): label evaluator select and notes textarea in invitation dialog").
+
+Open `dev` PR, no coverage yet (test-sync §9 sweep, 2026-10-08). Gated with
+`pending_pr` — skipped in CI until GitHub reports #443 merged.
+
+PR #443 head: db9cc7e89c43f437c18e7e7b565a486ab60fd598 (as of 2026-10-08).
+If review drags on, re-check the head hasn't moved before this merges.
+
+What changed: in the "Invite Evaluator" dialog (AuditorInvitation.tsx,
+opened from an AI model's version row via "Invite Evaluators"), the
+"Select Evaluator" <select> and the "Notes (optional)" <textarea> each had
+a <label> rendered visually above them but with no `htmlFor`/`id` link, so
+screen readers announced both controls with no name. #443 adds the
+`htmlFor`/`id` pair to both.
+
+Proof approach: Playwright's `get_by_label()` resolves a control through
+the real accessible-name algorithm (the same one assistive tech uses) --
+it does NOT just look for nearby text. Before the fix, with no
+`htmlFor`/`id`, `get_by_label("Select Evaluator")` and
+`get_by_label("Notes (optional)")` resolve to zero elements even though
+the label text is visibly right above the control. After the fix, each
+resolves to exactly one element. That is a genuine behaviour assertion,
+not a cosmetically-flipped expected value.
+
+UNVERIFIED — not executed, no credentials in this cloud sandbox. The
+invitation dialog is behind the AI Maker dashboard (Keycloak login via
+`authenticated_page_fast`), which this environment cannot reach, and it also
+requires a model that has at least one version with available evaluators
+to invite -- a data precondition that also needs a real login to check.
+A human with TEST_EMAIL_1/TEST_PASSWORD_1 must run this locally:
+
+    pytest tests/accessibility/test_auditor_invitation_a11y.py -v --tb=short -rs
+
+Red/green proof to perform, since #443 isn't on dev yet:
+  1. Run against dev now -> expect FAIL: `get_by_label(...)` finds 0
+     elements for both the evaluator select and the notes textarea (old
+     markup has no htmlFor/id).
+  2. `gh pr checkout 443` (frontend repo) and serve locally against the dev
+     API per SKILL.md section 6's local-serve recipe (note: the local-serve
+     stub-auth caveat means the header nav won't render, so reaching this
+     page may need to go directly to the model detail URL rather than via
+     nav click) -> expect PASS: both resolve to exactly one element.
+  3. Restore pointing at dev, confirm the gate (`pending_pr
+     ParakhAI-frontend#443: not merged yet` skip reason).
+"""
+
+import pytest
+from playwright.sync_api import Page
+
+from pages.models_page import ModelsPage
+
+pytestmark = [
+    pytest.mark.accessibility,
+    pytest.mark.auth,
+    pytest.mark.pending_pr("ParakhAI-frontend#443"),
+    pytest.mark.timeout(120),
+]
+
+
+@pytest.fixture
+def invite_dialog_page(authenticated_page_fast: Page) -> Page:
+    """Navigate to a model detail page and open the 'Invite Evaluator' dialog.
+
+    Skips (does not fail) when there's no model, no version, or no
+    available evaluators to invite — those are dev-data gaps, not what
+    this test is checking.
+    """
+    mp = ModelsPage(authenticated_page_fast)
+    mp.go_to_model_detail()
+    if not mp.is_invite_auditors_visible():
+        pytest.skip("No 'Invite Evaluators' link visible on dev — no model version to test")
+    if not mp.click_invite_auditors():
+        pytest.skip("'Invite Evaluator' dialog did not open")
+    return authenticated_page_fast
+
+
+class TestInvitationDialogAccessibleLabels:
+    """Select Evaluator and Notes must each resolve through their <label>."""
+
+    def test_evaluator_select_has_accessible_label(self, invite_dialog_page: Page):
+        if not invite_dialog_page.locator("select").first.is_visible():
+            pytest.skip(
+                "Evaluator <select> not rendered — likely 'no evaluators available' "
+                "state (loading list, or none invitable) rather than #443's target markup"
+            )
+        matches = invite_dialog_page.get_by_label("Select Evaluator")
+        assert matches.count() == 1, (
+            "'Select Evaluator' must resolve to exactly one control via its "
+            f"accessible label, found {matches.count()} (PR #443 adds the "
+            "missing htmlFor/id pair)"
+        )
+
+    def test_notes_textarea_has_accessible_label(self, invite_dialog_page: Page):
+        matches = invite_dialog_page.get_by_label("Notes (optional)")
+        assert matches.count() == 1, (
+            "'Notes (optional)' must resolve to exactly one control via its "
+            f"accessible label, found {matches.count()} (PR #443 adds the "
+            "missing htmlFor/id pair)"
+        )