diff --git a/app/harness/inference.py b/app/harness/inference.py
index 911dde5..42f58c9 100644
--- a/app/harness/inference.py
+++ b/app/harness/inference.py
@@ -388,6 +388,7 @@ async def _run_single_question(
     log: HarnessLog,
     max_steps: int,
     plugins: list[object],
+    run_id: str,
 ) -> dict[str, Any]:
     """执行单道题目的 Agent 推理。
 
@@ -402,6 +403,8 @@ async def _run_single_question(
         log: HarnessLog 实例（线程安全）。
         max_steps: AgentLoop 最大步数。
         plugins: pluggy 插件列表。
+        run_id: 运行标识，用作 cache_salt——run_id 含 _e{epoch} 天然跨 epoch 重采样、
+            同 epoch 续跑命中缓存（算法 #10 透传）。
 
     返回:
         预测结果字典（含 video_id, question_id, prediction, answer 等）。
@@ -431,6 +434,7 @@ async def _run_single_question(
             dispatcher,
             plugins=plugins,
             session_id=qa.question_id,
+            cache_salt=run_id,
         )
 
         result_dict = loop_result.result if isinstance(loop_result.result, dict) else {}
@@ -558,6 +562,7 @@ async def run_inference(
                 log=log,
                 max_steps=max_steps,
                 plugins=plugins,
+                run_id=run_id,
             )
             logger.info(
                 "[{}/{}] {} QA {} 完成 (stop={})",
diff --git a/core/agent/loop.py b/core/agent/loop.py
index 405292d..059285d 100644
--- a/core/agent/loop.py
+++ b/core/agent/loop.py
@@ -108,6 +108,7 @@ class AgentLoop:
         plugins: list[object] | None = None,
         *,
         session_id: str | None = None,
+        cache_salt: str | None = None,
     ) -> LoopResult:
         """执行 Thinking+JSON 推理循环。
 
@@ -117,6 +118,7 @@ class AgentLoop:
             tool_dispatcher: 工具调度器，ToolDispatcher Protocol 实例。
             plugins: pluggy 插件列表。
             session_id: 会话 ID，透传给 LLMProvider。
+            cache_salt: 缓存盐，透传给 LLMProvider（如训练用 run_id 跨 epoch 重采样）。
 
         返回:
             LoopResult 实例，包含推理步骤、token 用量、终止原因。
@@ -138,7 +140,7 @@ class AgentLoop:
             # Phase 1: LLM 调用（步级重试：防穿透 GovernedLLMClient 的瞬时异常）
             try:
                 response = await self._call_llm_with_step_retry(
-                    messages, token_usage, session_id=session_id
+                    messages, token_usage, session_id=session_id, cache_salt=cache_salt
                 )
             except Exception as e:
                 logger.error("LLM API 调用失败（{}）: {}", type(e).__name__, e)
@@ -269,6 +271,7 @@ class AgentLoop:
         token_usage: dict[str, int],
         *,
         session_id: str | None = None,
+        cache_salt: str | None = None,
     ) -> LLMResponse:
         """带步级重试的 LLM 调用，兜底穿透治理层重试栈的瞬时异常。
 
@@ -292,7 +295,9 @@ class AgentLoop:
         step_attempt = 0
         while True:
             try:
-                return await self._call_llm(messages, token_usage, session_id=session_id)
+                return await self._call_llm(
+                    messages, token_usage, session_id=session_id, cache_salt=cache_salt
+                )
             except self._retryable_exceptions as e:
                 step_attempt += 1
                 if step_attempt > self._step_retries:
@@ -315,6 +320,7 @@ class AgentLoop:
         token_usage: dict[str, int],
         *,
         session_id: str | None = None,
+        cache_salt: str | None = None,
     ) -> LLMResponse:
         """调用 LLMProvider 并累加 token 使用量。
 
@@ -322,11 +328,12 @@ class AgentLoop:
             messages: 消息历史。
             token_usage: 可变字典，就地累加。
             session_id: 会话 ID，透传给 LLMProvider。
+            cache_salt: 缓存盐，透传给 LLMProvider（跨 epoch 重采样）。
 
         返回:
             LLMResponse 实例。
         """
-        response = await self._llm.chat(messages, session_id=session_id)
+        response = await self._llm.chat(messages, session_id=session_id, cache_salt=cache_salt)
         token_usage["prompt_tokens"] += response.prompt_tokens
         token_usage["completion_tokens"] += response.completion_tokens
         return response
diff --git a/tests/unit/test_agent_loop.py b/tests/unit/test_agent_loop.py
index 3ca244b..2ce7383 100644
--- a/tests/unit/test_agent_loop.py
+++ b/tests/unit/test_agent_loop.py
@@ -94,6 +94,17 @@ def _invalid_tool_json() -> str:
 class TestAgentLoop:
     """AgentLoop 推理循环引擎测试。"""
 
+    @pytest.mark.asyncio
+    async def test_forwards_cache_salt(self) -> None:
+        """AgentLoop.run(cache_salt=...) 透传到 llm.chat（算法 #10 跨 epoch 重采样）。"""
+        llm = AsyncMock()
+        llm.chat.return_value = _make_response(_submit_json())
+
+        loop = AgentLoop(llm=llm, max_steps=10)
+        await loop.run("system", "user", _StubDispatcher(), cache_salt="run:e2")
+
+        assert llm.chat.call_args.kwargs["cache_salt"] == "run:e2"
+
     @pytest.mark.asyncio
     async def test_submit_answer_terminates_loop(self) -> None:
         """submit_answer 终止循环 → finished, result=args, steps_used=1。"""
diff --git a/tests/unit/test_harness_inference.py b/tests/unit/test_harness_inference.py
index fd57b04..5481cb9 100644
--- a/tests/unit/test_harness_inference.py
+++ b/tests/unit/test_harness_inference.py
@@ -698,6 +698,7 @@ class TestConcurrencyControl:
             *,
             session_id: str | None = None,
             parent_call_id: str | None = None,
+            cache_salt: str | None = None,
         ) -> LLMResponse:
             nonlocal current_concurrent, max_concurrent
             current_concurrent += 1