diff --git a/fam-edge/src/fam_edge/person_service.py b/fam-edge/src/fam_edge/person_service.py
index 77b0b08..7457ab3 100644
--- a/fam-edge/src/fam_edge/person_service.py
+++ b/fam-edge/src/fam_edge/person_service.py
@@ -25,6 +25,34 @@ from . import oracle_db
 
 logger = setup_logger('fam-edge.person_service')
 
+# 年龄档分桶：未成年与成年是稳定特征，跨桶绝不可能是同一个人
+_AGE_CHILD = {'幼儿', '儿童', '婴儿', '小孩', '少年', '青少年'}
+_AGE_ADULT = {'青年', '中年', '老年', '成年'}
+
+
+def _age_bucket(v: str) -> str:
+    if v in _AGE_CHILD:
+        return 'child'
+    if v in _AGE_ADULT:
+        return 'adult'
+    return ''
+
+
+def _features_conflict(fa: dict, fb: dict) -> str:
+    """硬冲突检测：性别不同 / 年龄档跨未成年-成年的两个人绝不合并。
+
+    LLM 合并只作为提议，落库前必须通过本校验（弱模型会把 爷爷/宝宝 全并进
+    人物A 这类错误）。返回冲突原因，无冲突返回空串。
+    """
+    ga, gb = str(fa.get('gender', '')), str(fb.get('gender', ''))
+    if ga and gb and 'unknown' not in (ga.lower(), gb.lower()) and ga != gb:
+        return f'性别冲突 {ga} vs {gb}'
+    ba = _age_bucket(str(fa.get('age_band', '')))
+    bb = _age_bucket(str(fb.get('age_band', '')))
+    if ba and bb and ba != bb:
+        return f'年龄档冲突 {fa.get("age_band")} vs {fb.get("age_band")}'
+    return ''
+
 
 class PersonService:
     def __init__(self, db: oracle_db.OracleDB):
@@ -82,19 +110,37 @@ class PersonService:
 
         # 4. 落库：canonical 若是另一个 label（target），解析为其已有 canonical，保证同一身份统一显示名
         label_to_canonical = {r['label']: (r['canonical_name'] or r['label']) for r in rows}
-        updated = 0
+
+        def _parse_feats(js):
+            try:
+                f = json.loads(js or '{}')
+                return f if isinstance(f, dict) else {}
+            except (ValueError, TypeError):
+                return {}
+
+        feat_map = {r['label']: _parse_feats(r['features_json']) for r in rows}
+        updated = blocked = 0
         for label, canonical in mapping.items():
             if label in manual:
                 continue  # 手动命名优先
             if canonical and canonical != label:
                 resolved = label_to_canonical.get(canonical, canonical)
+                conflict = _features_conflict(
+                    feat_map.get(label, {}), feat_map.get(resolved, {}))
+                if conflict:
+                    # 硬规则否决 LLM 提议：性别/年龄档冲突的绝不可能是同一人
+                    logger.warning(
+                        f"PersonService: 否决 LLM 合并 {label} -> {resolved}（{conflict}）")
+                    blocked += 1
+                    continue
                 self.db.set_canonical(label, resolved, source='llm')
                 updated += 1
         self.db.record_activity(
             'person', 'merge_done',
             f"校准 {len(label_videos)} 标签 + 聚合 {len(uid_features)} 特征，"
-            f"LLM 合并更新 {updated} 条（{', '.join(list(mapping)[:6])}）")
-        logger.info(f"PersonService: LLM 合并完成，更新 {updated} 条")
+            f"LLM 合并更新 {updated} 条 / 硬规则否决 {blocked} 条"
+            f"（{', '.join(list(mapping)[:6])}）")
+        logger.info(f"PersonService: LLM 合并完成，更新 {updated} 条，硬规则否决 {blocked} 条")
 
     def _aggregate_features(self) -> Dict[str, Dict]:
         """从 events.person_appearances_json 聚合每个 uid 的合并特征。