diff --git a/services/api/workers/resolve.py b/services/api/workers/resolve.py
index b2e5b5e..8090ad5 100644
--- a/services/api/workers/resolve.py
+++ b/services/api/workers/resolve.py
@@ -1,9 +1,13 @@
 """Thin adapter from the Python pipeline to 0pt's canonical TS ER engine.
 
 Python no longer scores identity. It sends opaque records to
-/api/0pt/internal/er/candidates, then only applies the returned canonical
-`match` decisions. If the canonical engine is unavailable, records are
-returned UNMERGED (fail closed against false identity merges).
+/api/0pt/internal/er/candidates, then applies only explicit hard-rule `match`
+decisions. Score/fuzzy decisions are review evidence only and can never merge
+records in this adapter, even if an older or compromised bridge returns
+`status=match, decided_by=score`.
+
+If the canonical engine is unavailable, records are returned UNMERGED (fail
+closed against false identity merges).
 
 A canonical hard non-match is also a cluster-wide constraint: transitive
 matches are never allowed to pull two records with conflicting stable identity
@@ -91,15 +95,18 @@ def _apply_matches(records: list[dict[str, Any]], pairs: list[dict[str, Any]]) -
         members.pop(drop, None)
         return True
 
+    # Defence in depth: only stable-identifier hard rules may create canonical
+    # identity links. Score/fuzzy "match" responses are ignored at the consumer
+    # boundary even if an upstream/legacy route emits one.
     match_pairs = [
         p for p in pairs
         if p.get("status") == "match"
+        and p.get("decided_by") not in (None, "score")
         and p.get("record_a") in known_ids
         and p.get("record_b") in known_ids
     ]
     match_pairs.sort(
         key=lambda p: (
-            0 if p.get("decided_by") not in (None, "score") else 1,
             -float(p.get("score") or 0.0),
             str(p.get("candidate_id") or ""),
         )