diff --git a/tests/graph/test_fanout_race.py b/tests/graph/test_fanout_race.py
new file mode 100644
index 0000000..bc53283
--- /dev/null
+++ b/tests/graph/test_fanout_race.py
@@ -0,0 +1,190 @@
+"""W2.B.4 — fanout reducer race test.
+
+Per BUILD_PLAN W2.B.4.a: build a graph with 4 executor branches, each
+sleeping a randomized 0-500ms, and assert the final merged state
+contains all 4 outputs in deterministic order regardless of who
+finished first.
+
+This is the critical invariant the HMAC ledger chain depends on: if two
+runs of the same case produce different `executor_results` orderings,
+they hash to different bytes and the chain integrity check fails. The
+reducer in `verdict/graph/reducers.py` sorts by
+`(branch_name, sequence_id)` so racing branches always merge to a
+byte-identical list.
+
+The race is exercised against a **real LangGraph fanout** (no mocks),
+using LangGraph's `Send` API to dispatch the 4 parallel branches.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import random
+import time
+from typing import Annotated, Any
+
+from langgraph.graph import END, START, StateGraph
+from langgraph.types import Send
+from typing_extensions import TypedDict
+
+from verdict.graph.reducers import append_executor_results
+
+BRANCH_NAMES = ("vol_exec", "hay_exec", "pls_exec", "mft_exec")
+
+
+class _RaceState(TypedDict, total=False):
+    """Minimal state for the race test — only the reducer-annotated
+    list is exercised."""
+
+    pending_branches: list[str]
+    executor_results: Annotated[list[dict[str, Any]], append_executor_results]
+
+
+# ---------------------------------------------------------------------------
+# Race graph: dispatcher -> 4 parallel `branch` workers (Send fanout) -> END.
+# ---------------------------------------------------------------------------
+
+
+def _dispatch(state: _RaceState) -> list[Send]:
+    """Conditional-edge router that fans out to N parallel `branch`
+    invocations via Send."""
+    return [Send("branch", {"branch_name": name}) for name in state["pending_branches"]]
+
+
+async def _branch(payload: dict[str, Any]) -> dict[str, Any]:
+    """One executor branch — sleeps a random 0-500ms then emits a
+    single result dict.
+
+    The result includes `branch_name` and `sequence_id` so the reducer
+    can produce a deterministic order regardless of completion time.
+    """
+    branch_name = payload["branch_name"]
+    # Real wall-clock sleep — exercising the actual race window.
+    delay_ms = random.randint(0, 500)
+    await asyncio.sleep(delay_ms / 1000.0)
+    return {
+        "executor_results": [
+            {
+                "branch_name": branch_name,
+                "sequence_id": 0,
+                "completed_at_ns": time.monotonic_ns(),
+            }
+        ]
+    }
+
+
+def _build_race_graph():
+    g: StateGraph = StateGraph(_RaceState)
+    g.add_node("branch", _branch)
+    # `add_conditional_edges` with a `Send`-returning router is the
+    # documented LangGraph fanout idiom.
+    g.add_conditional_edges(START, _dispatch, ["branch"])
+    g.add_edge("branch", END)
+    return g.compile()
+
+
+# ---------------------------------------------------------------------------
+# Tests
+# ---------------------------------------------------------------------------
+
+
+async def test_4_executors_merge_deterministically() -> None:
+    """W2.B.4.a — 4 executors, randomized 0-500ms sleep each; assert
+    final state contains all 4 outputs in deterministic order."""
+    compiled = _build_race_graph()
+
+    # Run the same fanout TWICE. Each run randomizes sleeps fresh, so
+    # the wall-clock completion order will almost certainly differ.
+    # Both final `executor_results` orderings must still be byte-
+    # identical because the reducer sorts deterministically.
+    random.seed(1)
+    out_a = await compiled.ainvoke({"pending_branches": list(BRANCH_NAMES)})
+    random.seed(2)
+    out_b = await compiled.ainvoke({"pending_branches": list(BRANCH_NAMES)})
+
+    names_a = [r["branch_name"] for r in out_a["executor_results"]]
+    names_b = [r["branch_name"] for r in out_b["executor_results"]]
+
+    assert names_a == sorted(BRANCH_NAMES), (
+        f"merged order is not deterministic on run A: got {names_a}"
+    )
+    assert names_b == sorted(BRANCH_NAMES), (
+        f"merged order is not deterministic on run B: got {names_b}"
+    )
+    # The names list must be identical across runs even though the
+    # completion timestamps differ — that's the HMAC chain invariant.
+    assert names_a == names_b
+
+
+async def test_all_4_branches_present() -> None:
+    """No branch is dropped on the merge — partial results would
+    masquerade as a successful 4-way fanout and bypass the
+    ARCHITECTURE.md §1 empty-set rule.
+    """
+    compiled = _build_race_graph()
+    out = await compiled.ainvoke({"pending_branches": list(BRANCH_NAMES)})
+    assert len(out["executor_results"]) == len(BRANCH_NAMES)
+    assert {r["branch_name"] for r in out["executor_results"]} == set(BRANCH_NAMES)
+
+
+async def test_reducer_idempotent_on_double_invoke() -> None:
+    """Calling .ainvoke() twice on the same compiled graph with the
+    same input must yield the same merged result — the reducer cannot
+    accumulate cross-invocation state.
+    """
+    compiled = _build_race_graph()
+    random.seed(42)
+    out1 = await compiled.ainvoke({"pending_branches": list(BRANCH_NAMES)})
+    random.seed(7)  # different seed, different wall-clock order
+    out2 = await compiled.ainvoke({"pending_branches": list(BRANCH_NAMES)})
+    assert [r["branch_name"] for r in out1["executor_results"]] == [
+        r["branch_name"] for r in out2["executor_results"]
+    ]
+
+
+# ---------------------------------------------------------------------------
+# Pure-reducer unit tests (no LangGraph) — pin the contract independently.
+# ---------------------------------------------------------------------------
+
+
+def test_reducer_handles_none_left_and_right() -> None:
+    """First write (left=None) and no-update (right=None) must not crash."""
+    assert append_executor_results(None, None) == []
+    assert append_executor_results(None, [{"branch_name": "x", "sequence_id": 0}]) == [
+        {"branch_name": "x", "sequence_id": 0}
+    ]
+    assert append_executor_results([{"branch_name": "x", "sequence_id": 0}], None) == [
+        {"branch_name": "x", "sequence_id": 0}
+    ]
+
+
+def test_reducer_sorts_by_branch_then_sequence() -> None:
+    """Reducer deterministic order: (branch_name, sequence_id)."""
+    items = [
+        {"branch_name": "vol_exec", "sequence_id": 1},
+        {"branch_name": "hay_exec", "sequence_id": 0},
+        {"branch_name": "vol_exec", "sequence_id": 0},
+        {"branch_name": "mft_exec", "sequence_id": 0},
+    ]
+    merged = append_executor_results([], items)
+    assert [r["branch_name"] for r in merged] == [
+        "hay_exec", "mft_exec", "vol_exec", "vol_exec"
+    ]
+    # Within vol_exec, sequence_id 0 sorts before 1.
+    vol_seqs = [r["sequence_id"] for r in merged if r["branch_name"] == "vol_exec"]
+    assert vol_seqs == [0, 1]
+
+
+def test_reducer_keyless_items_sort_to_tail_in_insertion_order() -> None:
+    """Items missing branch_name keep stable insertion order at the
+    tail of the merged list."""
+    items = [
+        {"foo": "first"},          # no branch_name
+        {"branch_name": "vol_exec", "sequence_id": 0},
+        {"foo": "second"},         # no branch_name
+    ]
+    merged = append_executor_results([], items)
+    assert merged[0] == {"branch_name": "vol_exec", "sequence_id": 0}
+    # Stable insertion order at the tail.
+    assert merged[1] == {"foo": "first"}
+    assert merged[2] == {"foo": "second"}