diff --git a/backend/scripts/mdm_pass.py b/backend/scripts/mdm_pass.py
index 6a8abfb..046604b 100644
--- a/backend/scripts/mdm_pass.py
+++ b/backend/scripts/mdm_pass.py
@@ -29,6 +29,7 @@ import json
 import re
 import sys
 import uuid
+from collections import Counter
 from pathlib import Path
 
 from sqlalchemy.orm.attributes import flag_modified
@@ -105,6 +106,26 @@ def conforms(article) -> bool:
     return all(required in titles for required in REQUIRED)
 
 
+def _multiset_problem(what: str, before: list[str], after: list[str]) -> list[str]:
+    """Describe how two multisets differ, counting repeats.
+
+    Comparing sets alone reported "lost -, added -" when a token simply appeared
+    a different number of times — which reads as nothing having changed, on a
+    line explaining a refusal.
+    """
+    want, got = Counter(before), Counter(after)
+    if want == got:
+        return []
+    lost = {k: want[k] - got.get(k, 0) for k in want if want[k] > got.get(k, 0)}
+    gained = {k: got[k] - want.get(k, 0) for k in got if got[k] > want.get(k, 0)}
+    parts = []
+    if lost:
+        parts.append("lost " + ", ".join(f"{k}x{n}" for k, n in sorted(lost.items())[:6]))
+    if gained:
+        parts.append("duplicated " + ", ".join(f"{k}x{n}" for k, n in sorted(gained.items())[:6]))
+    return [f"{what} changed ({'; '.join(parts)})"]
+
+
 def validate(before: list[dict], after: list[dict]) -> list[str]:
     """Reasons this restructure must not be applied. Empty means it is safe."""
     problems = []
@@ -128,15 +149,10 @@ def validate(before: list[dict], after: list[dict]) -> list[str]:
     if not new_text.strip():
         return ["empty"]
 
-    want, got = sorted(MARKER.findall(old_text)), sorted(MARKER.findall(new_text))
-    if want != got:
-        problems.append(f"cross-references changed (lost {sorted(set(want) - set(got)) or '-'}, "
-                        f"added {sorted(set(got) - set(want)) or '-'})")
-    want_n = sorted(NUMBER.findall(facts_only(old_text)))
-    got_n = sorted(NUMBER.findall(facts_only(new_text)))
-    if want_n != got_n:
-        problems.append(f"numbers changed (lost {sorted(set(want_n) - set(got_n)) or '-'}, "
-                        f"added {sorted(set(got_n) - set(want_n)) or '-'})")
+    problems += _multiset_problem("cross-references", MARKER.findall(old_text),
+                                  MARKER.findall(new_text))
+    problems += _multiset_problem("numbers", NUMBER.findall(facts_only(old_text)),
+                                  NUMBER.findall(facts_only(new_text)))
     ratio = len(new_text) / max(len(old_text), 1)
     if not MIN_RATIO <= ratio <= MAX_RATIO:
         problems.append(f"length {ratio:.2f}x of the original")