diff --git a/apps/api/src/models/ai.py b/apps/api/src/models/ai.py
index a37e5f19..50e72d70 100644
--- a/apps/api/src/models/ai.py
+++ b/apps/api/src/models/ai.py
@@ -130,11 +130,13 @@ class OpenClawDecision(BaseModel):
     )
 
     # === 信心度與影響範圍 ===
+    # 2026-03-29 ogt: 移除預設值，強制 LLM 必須輸出真實信心分數
+    # 如果 LLM 沒有輸出 confidence，解析時會補 0.5 並標記為 COLLAB
     confidence: float = Field(
-        default=0.8,
+        ...,  # REQUIRED - 不允許預設值
         ge=0.0,
         le=1.0,
-        description="決策信心度 (0-1)",
+        description="決策信心度 (0-1) - LLM 必須計算",
     )
     affected_services: list[str] = Field(
         default_factory=list,
diff --git a/apps/api/src/services/openclaw.py b/apps/api/src/services/openclaw.py
index a022383a..3a03b101 100644
--- a/apps/api/src/services/openclaw.py
+++ b/apps/api/src/services/openclaw.py
@@ -996,6 +996,17 @@ class OpenClawService:
             if "suggested_action" not in data:
                 data["suggested_action"] = "NO_ACTION"
 
+            # Step 2.5: 2026-03-29 ogt - 強制 confidence 必須由 LLM 輸出
+            # 如果 LLM 沒有輸出 confidence，強制設為 0.5 並標記為 COLLAB
+            if "confidence" not in data or not isinstance(data["confidence"], (int, float)):
+                logger.warning(
+                    "llm_missing_confidence",
+                    raw_confidence=data.get("confidence"),
+                    forcing_collab=True,
+                )
+                data["confidence"] = 0.5  # 低信心分數
+                data["primary_responsibility"] = "COLLAB"  # 強制協作處理
+
             # Step 3: 使用 Pydantic 驗證 (會自動正規化 risk_level, data_impact 等)
             decision = OpenClawDecision(**data)