diff --git a/tests/bench_gate/test_vocab_bridge_uplift.py b/tests/bench_gate/test_vocab_bridge_uplift.py
index 7a4dd59bf..06f38a1bd 100644
--- a/tests/bench_gate/test_vocab_bridge_uplift.py
+++ b/tests/bench_gate/test_vocab_bridge_uplift.py
@@ -22,19 +22,26 @@
   {
     "id": "row-id",
     "query": "raw query string passed to retrieve",
+    "k": 10,                                      # optional, default 10
     "store_beliefs": [
         {"id": "b1", "content": "...", "anchors": ["text", "..."]},
         ...
     ],
-    "expected_canonicals": ["sqlite", "python", "..."]
+    "expected_canonicals": ["sqlite", "python", "..."],
+    "expected_top_k": ["b1", "b2", ...]           # optional; A2 ship gate
   }
 
 `store_beliefs[i].anchors` is optional; when present, each anchor
 string is added as an inbound edge from a synthetic citing belief
 to seed the bridge with anchor-source surface forms (#148 parity).
 `expected_canonicals` is the set of canonical-entity tokens that
-the bridge SHOULD append for this query; the test asserts at least
-one is present in the rewritten output.
+the bridge SHOULD append for this query; the precondition test
+asserts at least one is present in the rewritten output.
+`expected_top_k` is the ground-truth belief-id ranking; the strict
+A2 ship gate (``test_vocab_bridge_ship_gate_runner_present``) runs
+``retrieve_v2`` with the bridge OFF then ON and asserts NDCG@k goes
+strictly up. Rows without ``expected_top_k`` are skipped by the
+ship gate but still exercised by the precondition gate.
 """
 from __future__ import annotations
 
@@ -133,3 +140,40 @@ def test_bridge_appends_at_least_one_expected_canonical(
         f"{n_hits}/{n_rows} rows ({coverage:.1%}); harvest/rewrite "
         f"regression suspected (was the surface-form pipeline broken?)"
     )
+
+
+@pytest.mark.bench_gated
+def test_vocab_bridge_ship_gate_runner_present(
+    aelfrice_corpus_root: Path,
+) -> None:
+    """The full A2 NDCG@k ship gate runs from
+    ``tests.retrieve_uplift_runner.run_vocab_bridge_uplift``. This test
+    skips when the runner is absent or when no row carries
+    ``expected_top_k`` — the runner is the operator-side gate for
+    flipping ``use_vocab_bridge`` to default-on (#433 Phase 2).
+    """
+    rows = load_corpus_module(aelfrice_corpus_root, "vocab_bridge")
+    assert rows, "vocab_bridge corpus produced zero rows"
+
+    runner_mod = pytest.importorskip(
+        "tests.retrieve_uplift_runner",
+        reason=(
+            "vocab_bridge uplift runner not yet wired "
+            "(operator gate; spec § A2 — pending lab-side corpus + scorer)"
+        ),
+    )
+
+    if not any(r.get("expected_top_k") for r in rows):
+        pytest.skip(
+            "vocab_bridge corpus has no rows with expected_top_k; "
+            "annotate at least one row before this gate can fire"
+        )
+
+    results = runner_mod.run_vocab_bridge_uplift(rows)
+    assert results.uplift > 0, (
+        "use_vocab_bridge must show strictly positive NDCG@k uplift\n"
+        f"  ON={results.mean_ndcg_on:.4f} "
+        f"OFF={results.mean_ndcg_off:.4f} "
+        f"uplift={results.uplift:+.4f} "
+        f"n_rows={results.n_rows}"
+    )
diff --git a/tests/retrieve_uplift_runner.py b/tests/retrieve_uplift_runner.py
index 91daf33ae..0f67a811d 100644
--- a/tests/retrieve_uplift_runner.py
+++ b/tests/retrieve_uplift_runner.py
@@ -609,6 +609,180 @@ def run_query_strategy_uplift(
     )
 
 
+# ---------------------------------------------------------------------
+# Vocabulary bridge (#433) — use_vocab_bridge bench gate (spec § A2).
+#
+# Consumed by tests/bench_gate/test_vocab_bridge_uplift.py.
+# Row schema (`tests/corpus/v2_0/vocab_bridge/*.jsonl`) — extends the
+# precondition row schema with `expected_top_k` and optional `k`:
+#
+#   {
+#     "id": "row-id",
+#     "query": "raw query string",
+#     "k": 10,
+#     "store_beliefs": [
+#         {"id": "b1", "content": "...", "anchors": ["text", "..."]},
+#         ...
+#     ],
+#     "expected_canonicals": ["sqlite", ...],   # used by precondition
+#     "expected_top_k":      ["b1", "b2", ...]  # ground-truth ranking
+#   }
+#
+# `expected_canonicals` stays the precondition surface — it lets the
+# appends-at-least-one gate fire without naming specific belief ids.
+# `expected_top_k` is the strict-NDCG ground truth: the bridge wins
+# only if rewriting the query surfaces these ids in this order vs the
+# default-OFF baseline.
+# ---------------------------------------------------------------------
+
+_BENCH_TS = "2026-05-08T00:00:00Z"
+
+
+@dataclass(frozen=True)
+class VocabBridgeUplift:
+    """#433 result shape.
+
+    Field names mirror ``DocLinkerUpliftResults`` / ``QueryStrategyUplift``
+    so the bench-gate failure-message formatter reads
+    ``mean_ndcg_off`` / ``mean_ndcg_on`` / ``uplift`` without
+    per-runner branching.
+    """
+
+    n_rows: int
+    mean_ndcg_off: float
+    mean_ndcg_on: float
+
+    @property
+    def uplift(self) -> float:
+        return self.mean_ndcg_on - self.mean_ndcg_off
+
+
+def _seed_store_for_vocab_bridge(
+    store: MemoryStore, row: dict,  # type: ignore[type-arg]
+) -> None:
+    """Seed a store from the vocab_bridge row shape.
+
+    Row carries ``store_beliefs`` (not ``beliefs``) where each entry
+    optionally lists ``anchors`` — surface-form strings written as
+    incoming-edge ``anchor_text`` from synthetic citing beliefs. This
+    mirrors the precondition seeder in
+    ``tests/bench_gate/test_vocab_bridge_uplift.py`` so the runner
+    sees an identical store shape; the bridge harvest/rewrite path
+    then operates on the same anchor-text universe under both arms.
+    """
+    for entry in row.get("store_beliefs", []):
+        bid = str(entry["id"])
+        store.insert_belief(
+            Belief(
+                id=bid,
+                content=str(entry["content"]),
+                content_hash=f"corpus:{bid}",
+                alpha=float(entry.get("alpha", 1.0)),
+                beta=float(entry.get("beta", 1.0)),
+                type=entry.get("type", BELIEF_FACTUAL),
+                lock_level=LOCK_NONE,
+                locked_at=None,
+                demotion_pressure=0,
+                created_at=_BENCH_TS,
+                last_retrieved_at=None,
+                origin=ORIGIN_AGENT_INFERRED,
+            )
+        )
+    for entry in row.get("store_beliefs", []):
+        for j, anchor in enumerate(entry.get("anchors") or []):
+            citer_id = f"_a_{entry['id']}_{j}"
+            if not store.get_belief(citer_id):
+                store.insert_belief(
+                    Belief(
+                        id=citer_id,
+                        content="anchor citer",
+                        content_hash=f"corpus:{citer_id}",
+                        alpha=1.0,
+                        beta=1.0,
+                        type=BELIEF_FACTUAL,
+                        lock_level=LOCK_NONE,
+                        locked_at=None,
+                        demotion_pressure=0,
+                        created_at=_BENCH_TS,
+                        last_retrieved_at=None,
+                        origin=ORIGIN_AGENT_INFERRED,
+                    )
+                )
+            store.insert_edge(
+                Edge(
+                    src=citer_id,
+                    dst=str(entry["id"]),
+                    type="CITES",
+                    weight=1.0,
+                    anchor_text=str(anchor),
+                )
+            )
+
+
+def run_vocab_bridge_uplift(
+    rows: list[dict],  # type: ignore[type-arg]
+) -> VocabBridgeUplift:
+    """Spec § A2 vocabulary-bridge uplift driver.
+
+    Per row, runs ``retrieve_v2`` twice on fresh stores seeded from
+    ``store_beliefs`` (with anchors): once with
+    ``use_vocab_bridge=False`` (baseline) and once with ``=True``.
+    Same query, same store shape, same ``k``. NDCG@k is scored against
+    ``expected_top_k`` and averaged across rows.
+
+    Rows missing ``expected_top_k`` are skipped (the precondition gate
+    only needs ``expected_canonicals``). The shipped substrate is
+    default-OFF: the bridge appends canonical-entity tokens to the
+    query, never substitutes — so on a corpus where BM25 alone
+    already ranks the relevant beliefs above noise, ``uplift`` will
+    be ~0 and the strict ``> 0`` ship-gate assertion will fail. That
+    is the correct gate-broken signal until labelled rows expose
+    surface-form gaps the bridge can close.
+    """
+    scoreable = [r for r in rows if r.get("expected_top_k")]
+    n = len(scoreable)
+    if n == 0:
+        return VocabBridgeUplift(0, 0.0, 0.0)
+
+    off_total = 0.0
+    on_total = 0.0
+    with tempfile.TemporaryDirectory() as tmp:
+        tmp_root = Path(tmp)
+        for row in scoreable:
+            k = _default_k(row)
+            expected = list(row["expected_top_k"])
+
+            for bridge_on in (False, True):
+                _db_counter[0] += 1
+                db = (
+                    tmp_root
+                    / f"vb_{row['id']}_{int(bridge_on)}_{_db_counter[0]}.db"
+                )
+                store = MemoryStore(str(db))
+                try:
+                    _seed_store_for_vocab_bridge(store, row)
+                    result = retrieve_v2(
+                        store, row["query"],
+                        budget=DEFAULT_TOKEN_BUDGET,
+                        use_entity_index=False,
+                        use_vocab_bridge=bridge_on,
+                    )
+                    returned = [b.id for b in result.beliefs[:k]]
+                finally:
+                    store.close()
+                ndcg = ndcg_at_k(returned, expected, k)
+                if bridge_on:
+                    on_total += ndcg
+                else:
+                    off_total += ndcg
+
+    return VocabBridgeUplift(
+        n_rows=n,
+        mean_ndcg_off=off_total / n,
+        mean_ndcg_on=on_total / n,
+    )
+
+
 def _format_table(results: list[FlagUplift]) -> str:
     lines = [
         f"{'flag':<28} {'n':>4} {'NDCG_off':>10} {'NDCG_on':>10} {'uplift':>10}",
diff --git a/tests/test_retrieve_uplift_runner.py b/tests/test_retrieve_uplift_runner.py
index 246faefb1..7e7e6afd7 100644
--- a/tests/test_retrieve_uplift_runner.py
+++ b/tests/test_retrieve_uplift_runner.py
@@ -282,6 +282,73 @@ def test_query_strategy_uplift_lowercase_no_extremes_means_zero_uplift() -> None
     assert r.uplift == 0.0
 
 
+def test_vocab_bridge_uplift_empty_input() -> None:
+    from tests.retrieve_uplift_runner import run_vocab_bridge_uplift
+
+    r = run_vocab_bridge_uplift([])
+    assert r.n_rows == 0
+    assert r.mean_ndcg_off == 0.0
+    assert r.mean_ndcg_on == 0.0
+    assert r.uplift == 0.0
+
+
+def test_vocab_bridge_uplift_skips_rows_without_expected_top_k() -> None:
+    """Precondition rows (those carrying only ``expected_canonicals``)
+    have no ground-truth ranking, so they must be skipped — not silently
+    scored as zero, which would dilute the mean and falsely shrink the
+    uplift signal. ``n_rows`` reflects the scoreable subset."""
+    from tests.retrieve_uplift_runner import run_vocab_bridge_uplift
+
+    rows = [
+        {
+            "id": "vb-precondition-only",
+            "query": "alpha",
+            "store_beliefs": [{"id": "a", "content": "alpha alpha"}],
+            "expected_canonicals": ["alpha"],
+        },
+    ]
+    r = run_vocab_bridge_uplift(rows)
+    assert r.n_rows == 0
+    assert r.uplift == 0.0
+
+
+def test_vocab_bridge_uplift_runs_on_synthetic_row() -> None:
+    """OFF/ON arms run end-to-end without raising; metrics are bounded.
+
+    Real uplift on a synthetic row is unlikely (the bridge harvests
+    canonical-entity superpositions from corpus surface forms; a
+    two-belief store doesn't expose a meaningful canonical universe).
+    The contract under test is shape + bounded metric, not that the
+    rewrite influences ranking — the corpus uplift is the lab-side
+    gate."""
+    from tests.retrieve_uplift_runner import run_vocab_bridge_uplift
+
+    row = {
+        "id": "vb-test-001",
+        "query": "memory store",
+        "k": 3,
+        "store_beliefs": [
+            {
+                "id": "b1",
+                "content": "the memory store persists beliefs",
+                "anchors": ["MemoryStore"],
+            },
+            {"id": "b2", "content": "the configuration file lives at /etc"},
+            {
+                "id": "b3",
+                "content": "the memory store uses sqlite",
+                "anchors": ["sqlite"],
+            },
+        ],
+        "expected_canonicals": ["sqlite"],
+        "expected_top_k": ["b1", "b3"],
+    }
+    r = run_vocab_bridge_uplift([row])
+    assert r.n_rows == 1
+    assert 0.0 <= r.mean_ndcg_off <= 1.0
+    assert 0.0 <= r.mean_ndcg_on <= 1.0
+
+
 def test_run_per_flag_uplift_covers_all_flags() -> None:
     """Hypothesis: the harness reports one row per registered flag.
     Falsifiable if a flag is silently dropped."""