diff --git a/docs/design/topic3-challenge-review-proposal.md b/docs/design/topic3-challenge-review-proposal.md
index b98b3f6..ce3fc55 100644
--- a/docs/design/topic3-challenge-review-proposal.md
+++ b/docs/design/topic3-challenge-review-proposal.md
@@ -315,42 +315,46 @@ research 任务：working → [庞统确认] → review → done
 
 ```yaml
 # guardrails.yaml
+#
+# L1 check 用 assert 字段（Python 表达式，Daemon eval 执行）
+# L2 check 用 prompt 字段（传给 subagent 的检查指令）
+
 task_types:
   coding:
     output_guardrails:
       - name: file_exists
-        check: "output.files | length > 0"
+        assert: "len(output.get('files', [])) > 0"
         severity: blocking
-        layer: L1                    # Daemon 直接操作
+        layer: L1
       - name: json_valid
-        check: "output.json_schema_valid"
+        assert: "output.get('json_schema_valid', False) == True"
         severity: blocking
         layer: L1
       - name: artifacts_exist
-        check: "output.artifacts_paths all exist"
+        assert: "all(os.path.exists(p) for p in output.get('artifacts_paths', []))"
         severity: blocking
-        layer: L1                    # CLI 附加校验
-      - name: code_quality
-        check: "scope_declaration vs task.truths"
+        layer: L1
+      - name: scope_alignment
+        prompt: "Compare the agent's scope_declaration against task truths. Check: is every truth covered? Are there deviations not declared?"
         severity: warning
-        layer: L2                    # spawn sub 检查
+        layer: L2
     output_review:
       required: true
-      mode: single_reviewer          # 单审
+      mode: single_reviewer
       max_rounds: 3
 
   deploy:
     plan_review:
       required: true
-      mode: debate                   # 对抗辩论
+      mode: debate
       max_rounds: 5
     output_guardrails:
       - name: no_direct_production
-        check: "output.target_env != 'production'"
-        severity: tripwire           # 立即中断
+        assert: "output.get('target_env') != 'production'"
+        severity: tripwire
         layer: L1
       - name: rollback_plan_exists
-        check: "output.rollback_plan != null"
+        assert: "output.get('rollback_plan') is not None"
         severity: blocking
         layer: L1
     output_review:
@@ -361,11 +365,11 @@ task_types:
   data:
     output_guardrails:
       - name: format_check
-        check: "output.format in ['csv', 'parquet', 'json']"
+        assert: "output.get('format') in ['csv', 'parquet', 'json']"
         severity: blocking
         layer: L1
     output_review:
-      required: false                # 低风险，Guardrail 自动检查
+      required: false
 
   research:
     output_review:
@@ -523,7 +527,7 @@ adversarial_instructions:
 
 ```yaml
 output_schema: "schemas/review-output.schema.json"
-verdict_options: ["approved", "rejected", "needs_revision", "approved_with_reservations"]
+verdict_options: ["approved", "rejected", "needs_revision"]
 
 required_fields:
   - verdict