diff --git a/internal/ownmanifest/seats_test.go b/internal/ownmanifest/seats_test.go
index e87d0466..9542e948 100644
--- a/internal/ownmanifest/seats_test.go
+++ b/internal/ownmanifest/seats_test.go
@@ -19,16 +19,25 @@ func TestEveryManifestInUseClaimsASeatTheMeshDefines(t *testing.T) {
 	if len(paths) == 0 {
 		t.Fatal("no manifests in the catalogue, so this proved nothing")
 	}
+	root := beside.Dir(t, "mesh-catalog")
 	paths = append(paths, "../../module.json")
 	var checked int
 	for _, path := range paths {
-		raw, err := os.ReadFile(path)
-		if err != nil {
-			t.Fatal(err)
-		}
 		// The SDK's claimProblems is unexported; the parser runs it on every manifest whose claims are well formed
-		// (go/catalogue manifest.go), so a manifest that parses claims only seats the mesh defines or declares.
-		m, err := catalogue.ParseManifest(raw)
+		// (go/catalogue manifest.go), so a manifest that parses claims only seats the mesh defines or declares. A
+		// catalogue manifest is judged as the merge check judges it, against the predefined list at the root of the tree it
+		// is read in (ADR 0348 §7, issue 914); this repository's own is judged alone.
+		var m catalogue.Manifest
+		var err error
+		if path == "../../module.json" {
+			var raw []byte
+			if raw, err = os.ReadFile(path); err != nil {
+				t.Fatal(err)
+			}
+			m, err = catalogue.ParseManifest(raw)
+		} else {
+			m, err = catalogue.ReadTreeManifest(root, filepath.Dir(path))
+		}
 		if err != nil {
 			t.Errorf("%s: %v", path, err)
 			continue
diff --git a/merge-check.sh b/merge-check.sh
index 9cd42198..4c5f0a52 100755
--- a/merge-check.sh
+++ b/merge-check.sh
@@ -28,8 +28,8 @@ export GOFLAGS=-mod=vendor GOPROXY=off
 
 # **The unit level's bound, in seconds** (issue 895): the slowest package's own test time of the untagged run, the
 # compile not counted. A run over it fails the check and names the slowest tests. It is 5 s since the tests carry the
-# `integration` tag (design 47 milestone 3): the slowest package, cmd/mesh-controller (internal/catalogue went to mesh-sdk, issue
-# 1381), measured 2.6-2.8 s on a quiet machine in about 2000 small tests, and the build node runs the store tests beside this run
+# `integration` tag (design 47 milestone 3): the slowest packages, cmd/mesh-controller and internal/catalogue, measured
+# 2.6-2.8 s on a quiet machine in about 2000 small tests each, and the build node runs the store tests beside this run
 # and other work besides, so twice the measure. How-we-build section 5 asks 1 s for the whole unit suite: issue 1055
 # (a catalogue loaded once per package run, fewer source-scanning tests, the timed run started alone).
 unit_bound=5