diff --git a/app/verify/signals.py b/app/verify/signals.py
index 00b8316..0ccfb89 100644
--- a/app/verify/signals.py
+++ b/app/verify/signals.py
@@ -110,8 +110,15 @@ def cpu_signals(rec: dict[str, Any], now_year: int) -> list[Signal]:
         _cmp_ge("threads_ge_cores", rec.get("threads"), rec.get("cores"), hard=True),
         _cmp_ge("boost_ge_base", rec.get("boost_clock_ghz"), rec.get("base_clock_ghz"), hard=True),
         _cmp_ge("max_tdp_ge_tdp", rec.get("max_tdp_w"), rec.get("tdp_w"), hard=False),
-        _cmp_ge("passmark_multi_ge_single", rec.get("passmark_cpu_mark"),
-                rec.get("passmark_single"), hard=False),
+        # No passmark_cpu_mark vs passmark_single check: PassMark's CPU Mark and
+        # its Single Thread Rating are separately normalised scales, so their
+        # magnitudes are not comparable. The old check read that as a defect —
+        # it failed 51 of 51 single-core parts, and its failures tracked absolute
+        # weakness rather than parallelism (failing median CPU Mark 641 vs 19,657
+        # for passing), with near-ties at the boundary (Core 2 Duo E8600: 1378 vs
+        # 1388). It only ever "held" because modern CPU Marks are large.
+        # Cinebench and Geekbench below DO report both figures on one scale, so
+        # multi >= single is a real expectation there.
         _cmp_ge("cb23_multi_ge_single", rec.get("cinebench_r23_multi"),
                 rec.get("cinebench_r23_single"), hard=False),
         _cmp_ge("gb_multi_ge_single", rec.get("geekbench_multi"),
diff --git a/tests/verify/test_signals.py b/tests/verify/test_signals.py
index f20bd7a..3a4138c 100644
--- a/tests/verify/test_signals.py
+++ b/tests/verify/test_signals.py
@@ -122,3 +122,17 @@ def test_core_count_under_the_wrong_vendor_field_fails():
 def test_core_count_in_the_right_field_passes():
     rec = {"manufacturer": "amd", "stream_processors": 2048, "release_date": "2020-01-01"}
     assert _named(signals.gpu_signals(rec, NOW), "vendor_core_field").result == "pass"
+
+
+def test_passmark_figures_are_not_compared():
+    """CPU Mark and Single Thread Rating are separately normalised PassMark
+    scales; a single-core part legitimately scores lower on the first."""
+    rec = {"cores": 1, "threads": 1, "passmark_cpu_mark": 195, "passmark_single": 301}
+    names = {s.name for s in signals.cpu_signals(rec, NOW)}
+    assert "passmark_multi_ge_single" not in names
+
+
+def test_cinebench_multi_below_single_still_fails():
+    """One scale, so multi < single there really is a defect."""
+    rec = {"cores": 8, "threads": 16, "cinebench_r23_multi": 900, "cinebench_r23_single": 1500}
+    assert _named(signals.cpu_signals(rec, NOW), "cb23_multi_ge_single").result == "fail"