diff --git a/fam-edge/src/fam_edge/ai_orchestrator/json_parser.py b/fam-edge/src/fam_edge/ai_orchestrator/json_parser.py
index ae0e7a8..c0b96c6 100644
--- a/fam-edge/src/fam_edge/ai_orchestrator/json_parser.py
+++ b/fam-edge/src/fam_edge/ai_orchestrator/json_parser.py
@@ -46,8 +46,12 @@ def parse_vlm_json(raw: str) -> dict:
 
 
 def validate_schema(data: dict) -> dict:
-    """Schema 校验 + 脏数据清洗"""
-    required = ["global_summary", "entities_json", "frame_details", "compute_provider"]
+    """Schema 校验 + 脏数据清洗
+
+    注意: compute_provider 是内部记账字段（orchestrator 在解析成功后填充），
+    提示词从不要求模型输出它，故不做必填校验。
+    """
+    required = ["global_summary", "entities_json", "frame_details"]
     for k in required:
         if k not in data:
             raise VLMOutputInvalidError(f"缺失字段: {k}")
@@ -94,11 +98,11 @@ def validate_schema(data: dict) -> dict:
         })
     data["frame_details"] = cleaned_frames
 
-    # compute_provider 校验为数组
-    if not isinstance(data["compute_provider"], list):
-        raise VLMOutputInvalidError("compute_provider 必须为数组")
-    if len(data["compute_provider"]) == 0:
-        raise VLMOutputInvalidError("compute_provider 不能为空数组")
-    data["compute_provider"] = [str(p) for p in data["compute_provider"]]
+    # compute_provider: 模型通常不输出（提示词不要求），缺失/非法时归一为空数组，
+    # 由 orchestrator.format_cloud_result 用实际成功的 provider 覆盖填充
+    cp = data.get("compute_provider")
+    if not isinstance(cp, list):
+        cp = []
+    data["compute_provider"] = [str(p) for p in cp]
 
     return data