diff --git a/internal/b/blueprints.go b/internal/b/blueprints.go
index 6599cf04..365d419c 100644
--- a/internal/b/blueprints.go
+++ b/internal/b/blueprints.go
@@ -110,6 +110,9 @@ type Event struct {
 	// The prev events of the event if we want to override or falsify them.
 	// If it is left at nil, MustCreateEvent will populate it automatically based on the forward extremities.
 	PrevEvents interface{}
+
+	// If this is a redaction, the event that it redacts
+	Redacts string
 }
 
 func MustValidate(bp Blueprint) Blueprint {
diff --git a/internal/client/client.go b/internal/client/client.go
index 0ccd0388..9ad48486 100644
--- a/internal/client/client.go
+++ b/internal/client/client.go
@@ -248,7 +248,8 @@ func (c *CSAPI) MustSync(t *testing.T, syncReq SyncReq) (gjson.Result, string) {
 // In the unlikely event that you need ordering on your checks, call MustSyncUntil multiple times
 // with a single checker, and reuse the returned since token, as in the "Incremental sync" example.
 //
-// Will time out after CSAPI.SyncUntilTimeout. Returns the latest since token used.
+// Will time out after CSAPI.SyncUntilTimeout. Returns the `next_batch` token from the final
+// response.
 func (c *CSAPI) MustSyncUntil(t *testing.T, syncReq SyncReq, checks ...SyncCheckOpt) string {
 	t.Helper()
 	start := time.Now()
diff --git a/internal/federation/server.go b/internal/federation/server.go
index b9759fe1..02fb7de1 100644
--- a/internal/federation/server.go
+++ b/internal/federation/server.go
@@ -263,6 +263,7 @@ func (s *Server) MustCreateEvent(t *testing.T, room *ServerRoom, ev b.Event) *go
 		PrevEvents: prevEvents,
 		Unsigned:   unsigned,
 		AuthEvents: ev.AuthEvents,
+		Redacts:    ev.Redacts,
 	}
 	if eb.AuthEvents == nil {
 		var stateNeeded gomatrixserverlib.StateNeeded
diff --git a/tests/csapi/sync_test.go b/tests/csapi/sync_test.go
index faca368f..4f49bf74 100644
--- a/tests/csapi/sync_test.go
+++ b/tests/csapi/sync_test.go
@@ -10,6 +10,7 @@ import (
 
 	"github.com/matrix-org/complement/internal/b"
 	"github.com/matrix-org/complement/internal/client"
+	"github.com/matrix-org/complement/internal/federation"
 	"github.com/matrix-org/complement/runtime"
 )
 
@@ -251,6 +252,120 @@ func TestSync(t *testing.T) {
 			res, _ := alice.MustSync(t, client.SyncReq{Since: nextBatch})
 			usersInPresenceEvents(t, res.Get("presence"), []string{})
 		})
+
+		t.Run("sync should succeed even if the sync token points to a redaction of an unknown event", func(t *testing.T) {
+			// this is a regression test for https://github.com/matrix-org/synapse/issues/12864
+			//
+			// The idea here is that we need a sync token which points to a redaction
+			// for an event which doesn't exist. Such a redaction may not be served to
+			// the client. This can lead to server bugs when the server tries to fetch
+			// the event corresponding to the sync token.
+			//
+			// The C-S API does not permit us to generate such a redaction event, so
+			// we have to poke it in from a federated server.
+			//
+			// The situation is complicated further by the very fact that we
+			// cannot see the faulty redaction, and therefore cannot tell whether
+			// our sync token includes it or not. The normal trick here would be
+			// to send another (regular) event as a sentinel, and then if that sentinel
+			// is returned by /sync, we can be sure the faulty event has also been
+			// processed. However, that doesn't work here, because doing so will mean
+			// that the sync token points to the sentinel rather than the redaction,
+			// negating the whole point of the test.
+			//
+			// Instead, as a rough proxy, we send a sentinel in a *different* room.
+			// There is no guarantee that the target server will process the events
+			// in the order we send them, but in practice it seems to get close
+			// enough.
+
+			t.Parallel()
+
+			// alice creates two rooms, which charlie (on our test server) joins
+			srv := federation.NewServer(t, deployment,
+				federation.HandleKeyRequests(),
+				federation.HandleTransactionRequests(nil, nil),
+			)
+			cancel := srv.Listen()
+			defer cancel()
+
+			charlie := srv.UserID("charlie")
+
+			redactionRoomID := alice.CreateRoom(t, map[string]interface{}{"preset": "public_chat"})
+			redactionRoom := srv.MustJoinRoom(t, deployment, "hs1", redactionRoomID, charlie)
+
+			sentinelRoomID := alice.CreateRoom(t, map[string]interface{}{"preset": "public_chat"})
+			sentinelRoom := srv.MustJoinRoom(t, deployment, "hs1", sentinelRoomID, charlie)
+
+			// charlie creates a bogus redaction, which he sends out, followed by
+			// a good event - in another room - to act as a sentinel. It's not
+			// guaranteed, but hopefully if the sentinel is received, so was the
+			// redaction.
+			redactionEvent := srv.MustCreateEvent(t, redactionRoom, b.Event{
+				Type:    "m.room.redaction",
+				Sender:  charlie,
+				Content: map[string]interface{}{},
+				Redacts: "$12345",
+			})
+			redactionRoom.AddEvent(redactionEvent)
+			t.Logf("Created redaction event %s", redactionEvent.EventID())
+			srv.MustSendTransaction(t, deployment, "hs1", []json.RawMessage{redactionEvent.JSON()}, nil)
+
+			sentinelEvent := srv.MustCreateEvent(t, sentinelRoom, b.Event{
+				Type:    "m.room.test",
+				Sender:  charlie,
+				Content: map[string]interface{}{"body": "1234"},
+			})
+			sentinelRoom.AddEvent(sentinelEvent)
+			srv.MustSendTransaction(t, deployment, "hs1", []json.RawMessage{redactionEvent.JSON(), sentinelEvent.JSON()}, nil)
+
+			// wait for the sentinel to arrive
+			nextBatch := alice.MustSyncUntil(t, client.SyncReq{}, client.SyncTimelineHasEventID(sentinelRoomID, sentinelEvent.EventID()))
+
+			// charlie sends another batch of events to force a gappy sync.
+			// We have to send 11 events to force a gap, since we use a filter with a timeline limit of 10 events.
+			pdus := make([]json.RawMessage, 11)
+			var lastSentEventId string
+			for i := range pdus {
+				ev := srv.MustCreateEvent(t, redactionRoom, b.Event{
+					Type:    "m.room.message",
+					Sender:  charlie,
+					Content: map[string]interface{}{},
+				})
+				redactionRoom.AddEvent(ev)
+				pdus[i] = ev.JSON()
+				lastSentEventId = ev.EventID()
+			}
+			srv.MustSendTransaction(t, deployment, "hs1", pdus, nil)
+			t.Logf("Sent filler events, with final event %s", lastSentEventId)
+
+			// sync, starting from the same ?since each time, until the final message turns up.
+			// This is basically an inlining of MustSyncUntil, with the key difference that we
+			// keep the same ?since each time, instead of incrementally syncing on each pass.
+			numResponsesReturned := 0
+			start := time.Now()
+			for {
+				if time.Since(start) > alice.SyncUntilTimeout {
+					t.Fatalf("%s: timed out after %v. Seen %d /sync responses", alice.UserID, time.Since(start), numResponsesReturned)
+				}
+				// sync, using a filter with a limit smaller than the number of PDUs we sent.
+				syncResponse, _ := alice.MustSync(t, client.SyncReq{Filter: filterID, Since: nextBatch})
+				numResponsesReturned += 1
+				timeline := syncResponse.Get("rooms.join." + client.GjsonEscape(redactionRoomID) + ".timeline")
+				timelineEvents := timeline.Get("events").Array()
+				lastEventIdInSync := timelineEvents[len(timelineEvents)-1].Get("event_id").String()
+
+				t.Logf("Iteration %d: /sync returned %d events, with final event %s", numResponsesReturned, len(timelineEvents), lastEventIdInSync)
+				if lastEventIdInSync == lastSentEventId {
+					// check we actually got a gappy sync - else this test isn't testing the right thing
+					if !timeline.Get("limited").Bool() {
+						t.Fatalf("Not a gappy sync after redaction")
+					}
+					break
+				}
+			}
+
+			// that's it - we successfully did a gappy sync.
+		})
 	})
 }