diff --git a/docs/architecture/rfcs/hierarchical-agent-stride-control-v0.md b/docs/architecture/rfcs/hierarchical-agent-stride-control-v0.md
index 3f2a4e559e..e290d5629d 100644
--- a/docs/architecture/rfcs/hierarchical-agent-stride-control-v0.md
+++ b/docs/architecture/rfcs/hierarchical-agent-stride-control-v0.md
@@ -404,6 +404,84 @@ protected-operation policy remain authoritative for permission. Supervisors,
 dashboards, and scheduler projections may recommend intervention but cannot
 silently convert a proposal into authority.
 
+### 7.5 Long-running work and decision checkpoints (proposed refinement)
+
+This 2026-10-04 refinement is a proposal for M3/M5, not shipped adaptive
+scheduling. It separates an **experiment boundary**, a **model decision
+checkpoint**, and an **accountable Turn boundary**. One experiment can span
+several Turns; every admitted accountable Turn still validates, writes back,
+and settles its own identity. Waiting for the experiment does not defer an
+owed debit, keep a Turn open indefinitely, or make earlier work free.
+
+The nearest owning capability/provider must define, before launch:
+
+- the question, candidate revision, comparison and bounded postcondition;
+- the authorized job identity, resource/time limits and recovery owner;
+- which observable result can change the next decision, and how to validate
+  that result; a training checkpoint, evaluation or build is meaningful only
+  under this declared contract;
+- the last consumed result revision, next observation time/deadline and
+  stop/cancel behavior; elapsed time and log growth alone are not evidence.
+
+These are capability/provider data. The kernel consumes identity, readiness,
+freshness and existing effect/settlement facts; it does not interpret epochs,
+benchmark scores, proof counts or process log prose. No generic job scheduler,
+worker launcher or new built-in capability is needed for this proposal.
+
+The proposed handoff is:
+
+| Phase | Owner and required behavior |
+|---|---|
+| Start bounded work | A normal selected Todo/Turn admits the effect. The provider persists job/candidate identity and an operational receipt. Settle that Turn truthfully: accepted launch proves launch, not the final experiment outcome. |
+| Observe while waiting | The existing host/runtime or capability observer checks the exact job without a reasoning-model call. Reuse one due monitor or existing observation channel; do not create a Todo per poll. Record observer CPU/IO cost even when no agent slot is spent. |
+| Return a decision checkpoint | The provider validates a new result against the declared postcondition, binds its revision/artifact digest and emits a compact observation. Failure, cancellation and deadline expiry must also produce actionable observations. |
+| Admit a decision | The existing scheduler/Turn path rechecks Goal lifetime, ownership, quota, capabilities and authority. A ready observation is input, not execution permission. Other runnable work can proceed while this experiment waits. |
+| Consume and continue | In the accountable consuming Turn, commit result adoption and the next step through the normal writeback before acknowledging consumption. Replayed/duplicate delivery recovers that receipt without duplicating launch or adoption. Settlement replay is idempotent for the same Turn identity; a newly admitted recovery Turn still settles normally. |
+
+Current Todo resume conditions include `monitor_changed` and `resume_at`; an
+implementation should use the existing typed monitor generation and timeout
+fallback where sufficient. **The missing integration is the provider-result
+binding and material-generation update**, not a new arbitrary `job_done` token.
+Only a validated decision checkpoint may advance that generation. The monitor
+must name the exact job/result source and bound Todo; unrelated log changes
+cannot release it. A safety observation deadline catches a lost notification
+or crashed process; it is not a periodic request to reinterpret unchanged work.
+Unsupported hosts retain current execution and show unavailable readiness,
+rather than silently claiming model-wake suppression.
+
+Identity must survive host/Agent restart: Goal lifetime, owning Agent/Todo,
+provider/run id, attempt or job generation, candidate revision, result revision
+and evidence/artifact refs. Keep raw logs and native job handles with the
+provider; expose only bounded authorized references. A PID alone is not a
+recoverable identity. Provider absence or an ambiguous external effect remains
+unknown under existing recovery rules; it cannot justify an automatic relaunch.
+The consumed cursor advances only with validated durable adoption, so a crash
+after result delivery re-offers the result. Stopped/revoked Goals and expired
+claims fence late results; replacement attempts cannot consume old results.
+
+Persist the incumbent artifact reference and its qualifying result separately
+from the running candidate. The provider updates it only after a comparable
+valid result satisfies its declared constraints. Recovery verifies reference
+availability and content before reuse; a missing candidate artifact does not
+silently replace the incumbent or authorize another launch.
+
+The first experiment is a **shadow observation** over one real capability
+runtime: report unchanged polls, decision checkpoints, result-to-decision
+latency, native outcome and total model/observer cost. Then qualify opt-in
+model-admission suppression only for that supported provider/host combination.
+Do not change the generic heartbeat prompt or all scheduler defaults. Turning
+the policy off restores fixed admission while retaining job/evidence history;
+it neither kills work implicitly nor changes acceptance or accounting.
+
+Acceptance requires real controlled process execution and recovery, not just
+serialized events: duplicate and out-of-order results, changed candidates,
+restart, missed notification, deadline, cancellation, unknown effect, competing
+claim and owner stop. A packaged frontend journey must show the bound work,
+waiting reason, last/next observation, result and stale/unavailable recovery
+through existing task/capability surfaces. CLI readback alone is a partial
+slice. Scientific usefulness follows the independent qualification in
+[research decisions](research-exploration-control-plane-v0.md#115-evidence-conditioned-experiment-decisions-proposed-refinement).
+
 ## 8. Measurement Model
 
 The first implementation should measure before it controls.
@@ -652,6 +730,8 @@ policy framework.
   transitions;
 - validate recommendations against independent acceptance rules;
 - retain Todo and replan as the only execution authorities.
+- qualify the proposed §7.5 provider checkpoint/monitor binding in shadow;
+  distinguish model decisions from experiment completion and Turn settlement.
 
 ### M4: Authority-stride shadow recommendations
 
@@ -665,6 +745,8 @@ policy framework.
 - retain hard ceilings and rollback;
 - compare repeated results against the pinned fixed profile;
 - publish limitations and failure modes with any claimed improvement.
+- for long-running work, qualify §7.5 lost-result recovery and default-off
+  parity before suppressing model admission; observer cost remains measured.
 
 ## 13. Validation Criteria
 
diff --git a/docs/architecture/rfcs/hierarchical-agent-stride-control-v0.zh-CN.md b/docs/architecture/rfcs/hierarchical-agent-stride-control-v0.zh-CN.md
index 8486b28878..15997e658b 100644
--- a/docs/architecture/rfcs/hierarchical-agent-stride-control-v0.zh-CN.md
+++ b/docs/architecture/rfcs/hierarchical-agent-stride-control-v0.zh-CN.md
@@ -364,6 +364,75 @@ Goal 与 vision state 继续拥有 acceptance 权限。User gate 与 protected-o
 policy 继续拥有 permission 权限。Supervisor、dashboard 与 scheduler projection
 可以建议干预，但不能静默把 proposal 变成 authority。
 
+### 7.5 长作业与决策 checkpoint（修订提案）
+
+这项 2026-10-04 修订属于 M3/M5 提案，不代表 adaptive scheduling 已交付。
+需要分开**实验边界**、**模型决策 checkpoint**和**可记账 Turn 边界**。
+一个实验可以跨多个 Turn；每个已获准且需要记账的 Turn 仍以自己的 identity
+完成 validation、writeback 和 settlement。等待实验不会延后欠付 debit、让
+Turn 无限保持打开，或把此前工作变成免费。
+
+最近的 capability/provider owner 应在启动前声明：
+
+- 问题、candidate revision、比较对象和有界 postcondition；
+- 已授权的 job identity、资源/时间上限和 recovery owner；
+- 什么可观察结果会影响下一决策，以及如何验证；训练 checkpoint、评测或
+  构建只有在这份已声明的合同下才有决策意义；
+- 已消费 result revision、下次观测时间/deadline、stop/cancel 行为；仅经过
+  一段时间或日志增长不是证据。
+
+这些是 capability/provider 数据。Kernel 消费 identity、readiness、freshness
+及现有 effect/settlement 事实，不解释 epoch、benchmark 分数、证明数量或进程
+日志 prose。本提案不需要新增通用 job scheduler、worker launcher 或 built-in
+capability。
+
+拟议 handoff 如下：
+
+| 阶段 | Owner 与必需行为 |
+|---|---|
+| 启动有界作业 | 普通已选 Todo/Turn 准入 effect。Provider 持久化 job/candidate identity 与 operational receipt。该 Turn 如实结算：启动成功只证明启动，不证明实验最终结果。 |
+| 等待中的观测 | 现有 host/runtime 或 capability observer 无需 reasoning-model 调用即可检查精确 job。复用一个 due monitor 或已有 observation channel，不为每次 poll 创建 Todo。即使不消费 agent slot，也记录 observer CPU/IO 成本。 |
+| 返回决策 checkpoint | Provider 按已声明的 postcondition 验证新结果，绑定 revision/artifact digest 并返回紧凑 observation。失败、取消和 deadline 到期同样必须形成可操作 observation。 |
+| 准入决策 | 现有 scheduler/Turn 路径重新检查 Goal lifetime、ownership、quota、capability 与 authority。Ready observation 是输入，不是执行许可。等待此实验时仍可推进其他 runnable 工作。 |
+| 消费并继续 | 在负责消费的可记账 Turn 中，通过普通 writeback 提交 result adoption 与下一步骤，再确认消费。重放/重复交付恢复原 receipt，不重复 launch 或 adoption。同一 Turn identity 的结算重放幂等；新获准 recovery Turn 仍正常结算。 |
+
+当前 Todo resume condition 已有 `monitor_changed` 与 `resume_at`。若足够，
+实现应复用 typed monitor generation 与 timeout fallback。
+**缺失的是 provider-result binding 和 material-generation 更新集成**，不是
+新增任意 `job_done` token。只有通过验证的决策 checkpoint 才能推进 generation。
+Monitor 必须绑定精确 job/result source 与 Todo；无关日志变化不能释放等待。
+Safety observation deadline 用来发现丢失通知或崩溃进程，不是周期性要求模型
+重新解释未变工作。不支持的 host 保持现有执行并显示 readiness unavailable，
+不能暗称已抑制模型唤醒。
+
+Identity 必须跨 host/Agent restart 保留：Goal lifetime、owning Agent/Todo、
+provider/run id、attempt 或 job generation、candidate revision、result revision
+与 evidence/artifact ref。Raw log/native job handle 留在 provider，仅投影有界
+且授权的引用。单个 PID 不是可恢复 identity。Provider 缺失或外部 effect 状态
+不明确时，按现有 recovery 规则保留 unknown，不能据此自动 relaunch。只有
+validated durable adoption 才推进 consumed cursor，因此交付后崩溃会重新提供
+结果。Stopped/revoked Goal 和过期 claim 拦截 late result；替换 attempt 不能
+消费旧结果。
+
+Incumbent artifact reference 及其 qualifying result 与当前运行 candidate 分开
+持久化。Provider 仅在可比有效结果满足已声明 constraint 后更新 incumbent。
+Recovery 在复用前验证引用可用性与内容；candidate artifact 缺失不能静默
+替换 incumbent，也不授权再次 launch。
+
+先在一个真实 capability runtime 做 **shadow observation**：报告 unchanged poll、
+decision checkpoint、result-to-decision latency、原生结果和模型/observer 总成本。
+随后仅对该 provider/host 组合资格验证 opt-in model-admission suppression。
+不改变通用 heartbeat prompt 或所有 scheduler 默认行为。关闭策略恢复 fixed
+admission，保留 job/evidence history，不隐式杀作业，不改变 acceptance 或 accounting。
+
+验收必须覆盖真实受控进程执行与 recovery，不只是 event 序列化：重复与乱序
+结果、candidate 改变、restart、丢失通知、deadline、取消、unknown effect、竞争
+claim 和 owner stop。Packaged frontend 旅程通过已有 task/capability surface 显示
+绑定工作、等待原因、上次/下次观测、结果及 stale/unavailable 恢复；仅 CLI
+readback 属于 partial slice。科学效用遵循
+[研究决策](research-exploration-control-plane-v0.zh-CN.md#115-由证据驱动的实验决策修订提案)
+的独立资格验证。
+
 ## 8. 测量模型
 
 第一版实现应先测量，再控制。
@@ -597,6 +666,9 @@ executor、scheduler 或 generic policy framework。
 - 用独立 acceptance rule 验证建议；
 - Todo 与 replan 仍是唯一 execution authority。
 
+- 在 shadow 中验证 §7.5 的 provider checkpoint/monitor binding，区分模型
+  决策、实验完成与 Turn settlement。
+
 ### M4：Authority-stride shadow recommendation
 
 - 区分 report 与 authority-changing intervention；
@@ -610,6 +682,9 @@ executor、scheduler 或 generic policy framework。
 - 对 pinned fixed profile 做 repeated comparison；
 - 任何 improvement claim 都同时公开 limitation 与 failure mode。
 
+- 长作业抑制 model admission 前，验证 §7.5 的 lost-result recovery 与
+  default-off parity；observer 成本继续纳入测量。
+
 ## 13. 验证标准
 
 只有证明以下各项，研究计划才算合格：
diff --git a/docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.md b/docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.md
index 29f54f3b9b..671ec2e877 100644
--- a/docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.md
+++ b/docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.md
@@ -655,6 +655,16 @@ layer at a time:
 The study should estimate model- and work-class-specific response curves, not
 one global tool-call or Todo-count threshold. Wider is not automatically better.
 
+The proposed [long-running-work refinement](hierarchical-agent-stride-control-v0.md#75-long-running-work-and-decision-checkpoints-proposed-refinement)
+tests **decision timing**, not deferred Turn accounting. First capture one
+provider's validated result/checkpoint and bound monitor in shadow; only a
+qualified opt-in arm may suppress unchanged model admission. Keep native
+evaluation, job resources, deadlines and every owed Turn settlement identical.
+Measure observer CPU/IO, model tokens/calls, result-to-decision latency, missed
+results, duplicate execution/debits and native outcome per total budget. A
+missing callback must recover through the declared observation deadline.
+Neither a longer training run nor fewer bookkeeping calls alone qualifies it.
+
 ### 8.2 Evidence, stall detection, and semantic replan
 
 The core hypothesis is that a durable coverage ledger plus semantic progress
@@ -668,6 +678,18 @@ whether direction changes create progress. DeepSWE validates whether the same
 mechanism improves repository outcomes without reward-specific shortcuts. ALE
 tests transfer to heterogeneous professional workflows.
 
+The proposed [experiment decision path](research-exploration-control-plane-v0.md#115-evidence-conditioned-experiment-decisions-proposed-refinement)
+qualifies result adoption separately from cadence. Hold session continuity,
+planning, model and total budget fixed; compare delivered evidence alone with
+evidence-linked continuation/successor adoption, then combine mechanisms only
+after their independent comparisons. Include valid negative evidence,
+inconclusive/conflicting measurements, failed prerequisites, justified repeats,
+uncovered same-family probes and changed inputs reopening a retired scope.
+Assess whether the executed next action follows the scoped result, alongside
+native outcome and cost. Citation/schema compliance alone is insufficient;
+forced pivots that suppress useful experiments count as failures. These are
+proposed studies, not current C2/C4 evidence or permission to launch jobs.
+
 ### 8.3 Research exploration and composition
 
 The [Research Exploration Control Plane RFC](./research-exploration-control-plane-v0.md)
diff --git a/docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.zh-CN.md b/docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.zh-CN.md
index 957987a6b3..b031e19e61 100644
--- a/docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.zh-CN.md
+++ b/docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.zh-CN.md
@@ -552,6 +552,14 @@ non-material event mix；它不能证明 benchmark integrity，也不能复用
 研究应估计 model/work-class-specific response curve，而不是一个全局 tool-call 或 Todo-count
 threshold。更宽并不天然更好。
 
+拟议[长作业修订](hierarchical-agent-stride-control-v0.zh-CN.md#75-长作业与决策-checkpoint修订提案)
+测试**决策时机**，不延后 Turn 记账。先 shadow capture 一个 provider 的 validated
+result/checkpoint 和绑定 monitor；只有通过资格验证的 opt-in arm 才能抑制
+unchanged model admission。原生评测、job 资源、deadline 与所有欠付 Turn
+settlement 保持一致。测量 observer CPU/IO、模型 token/call、result-to-decision
+latency、漏结果、重复执行/debit 与总预算下原生结果。Callback 丢失必须按已声明
+observation deadline 恢复。训练更久或 bookkeeping call 更少，单独都不够资格。
+
 ### 8.2 Evidence、空转检测与 semantic replan
 
 核心假设是 durable coverage ledger 与 semantic progress observation 能阻止重复 maintenance
@@ -563,6 +571,15 @@ LHTB 是首要动态实验场，因为 partial reward 与 checkpoint 可以显
 DeepSWE 验证相同机制能否改善 repository outcome，且不依赖 reward-specific shortcut。
 ALE 验证它能否迁移到异构专业工作流。
 
+拟议[实验决策路径](research-exploration-control-plane-v0.zh-CN.md#115-由证据驱动的实验决策修订提案)
+把 result adoption 与 cadence 分别验证。固定 session continuity、planning、model
+和总预算，对比仅交付 evidence 与 evidence-linked continuation/successor adoption，
+独立比较成立后再组合机制。包含有效负证据、inconclusive/conflicting measurement、
+失败前置条件、有依据复验、未覆盖的同 family probe 与 input 改变后重开 scope。
+除原生结果和成本外，判断实际执行的 next action 是否遵循 scoped result。
+Citation/schema compliance 单独不够；压制有用实验的强制 pivot 算失败。
+这些是拟议 study，不是当前 C2/C4 evidence，也不授予 job launch 权限。
+
 ### 8.3 研究探索与组合
 
 [研究型探索控制面 RFC](./research-exploration-control-plane-v0.zh-CN.md)定义 typed research
diff --git a/docs/architecture/rfcs/loopx-overall-roadmap-v0.md b/docs/architecture/rfcs/loopx-overall-roadmap-v0.md
index bac4e0a661..4908390f4a 100644
--- a/docs/architecture/rfcs/loopx-overall-roadmap-v0.md
+++ b/docs/architecture/rfcs/loopx-overall-roadmap-v0.md
@@ -329,8 +329,8 @@ subsystem was not performed. Section 8 records the focused audit.
 | [Human Attention Wishlist v0](human-attention-wishlist-v0.md) | S5/S11 | Accepted; Held | P3: reopen only on repeated second real need; sidecar cannot alter gates/quota/scheduling |
 | [Human-confirmed domain operations (v0)](human-confirmed-domain-operations-v0.md) | S8/S9 + R2/R3 | Accepted; canonical operation seam exists, managed native transport under owner review | Qualify source-context vs admitted-executor separation, exact human approval→one-shot consumption→domain evidence→original-route return; owned Turn/delegation need not await Desktop authentication. Live approval/effect/wakeup remain unqualified; finance provider stays separate |
 | [Provider-side authorization at effect acceptance (v0)](provider-effect-acceptance-v0.md) | S8/S9, supporting S2/S4 | Accepted design only; no runtime integration or qualified provider | M1: controlled provider and deterministic revoke/crash/replay conformance; strict production wiring remains gated by exact Goal lifetime, receipt retention and independent provider qualification |
-| [Research Exploration Control Plane v0](research-exploration-control-plane-v0.md) | S11/S3 | Accepted; partial M2 composition/successor | P1: independently verify observation/write-time gate/closure basis; defer inferred triggers and model selection |
-| [Hierarchical Agent Stride Control v0](hierarchical-agent-stride-control-v0.md) | S11/S7 | Accepted; M1 read-only observation | P2: matched shadow stride experiment with costs/events; no direct production cadence change |
+| [Research Exploration Control Plane v0](research-exploration-control-plane-v0.md) | S11/S3 | Accepted; partial M2 composition/successor; §11.5 result/adoption refinement proposed | P1: one real provider result -> scoped interpretation -> adopted step/successor, exact lineage/shared write gate and packaged readback; qualify shadow before opt-in, retain composition gaps; defer inference and M4 portfolio ranking |
+| [Hierarchical Agent Stride Control v0](hierarchical-agent-stride-control-v0.md) | S11/S7 | Accepted; M1 read-only observation; §7.5 checkpoint/monitor binding proposed | P2: matched shadow decision timing with observer/model costs and lost-result recovery; preserve per-Turn settlement, no default cadence change |
 | [Goal-scoped Capability Portfolio v0](goal-scoped-capability-portfolio-v0.md) | S1/S3/S6/S8/S11 | Accepted; read-only settings/context inspection slice | P1: correction→fresh-session decision across existing owners, then demand-driven composition and measured method evolution; no universal memory database or second opt-in |
 | [Post-Outcome Memory Utility Attribution v0](post-outcome-memory-utility-attribution-v0.md) | S6/S11 | Accepted; Stage 1 verified-outcome binding | P1 read-only reducer→P2 pilot: distinguish recalled/applied/utility, no automatic ranking change |
 | [Obelisk Session Evidence Provider v0](obelisk-session-evidence-provider-v0.md) | S6/S8 | Accepted; optional read-only evaluation | P2: explicit gap recall with resolvable provenance/access, off-path parity; no work authority |
diff --git a/docs/architecture/rfcs/loopx-overall-roadmap-v0.zh-CN.md b/docs/architecture/rfcs/loopx-overall-roadmap-v0.zh-CN.md
index cd3a62ef1e..d512d26a59 100644
--- a/docs/architecture/rfcs/loopx-overall-roadmap-v0.zh-CN.md
+++ b/docs/architecture/rfcs/loopx-overall-roadmap-v0.zh-CN.md
@@ -266,8 +266,8 @@ Muse 设计页在浏览器超时，其文章通过网页检索读取。本次调
 | [Human Attention Wishlist v0](human-attention-wishlist-v0.zh-CN.md) | S5/S11 | 已接受；Held | P3：第二个重复真实需求出现才重开；sidecar 不改变 gate/quota/调度 |
 | [Human-confirmed domain operations (v0)](human-confirmed-domain-operations-v0.zh-CN.md) | S8/S9 + R2/R3 | 已接受；规范操作接缝存在，受管原生传输待 owner review | 验收来源上下文与已准入执行者分离、精确用户批准→单次消费→垂域证据→原路返回；自有 Turn/delegation 不等待 Desktop 认证。真实批准/效果/唤醒仍未资格化；金融 provider 保持独立 |
 | [Provider 在效果接受点执行授权（v0）](provider-effect-acceptance-v0.zh-CN.md) | S8/S9，S2/S4 支撑 | 已接受设计；尚未接入 runtime，也未准入 provider | M1：controlled provider 与 deterministic revoke/crash/replay conformance；strict production 接线仍需精确 Goal 生命周期、receipt retention 与独立 provider 资格 |
-| [Research Exploration Control Plane v0](research-exploration-control-plane-v0.zh-CN.md) | S11/S3 | 已接受；M2 composition/successor 局部实现 | P1：observation/write-time gate/closure basis 独立验证；自选模型和推断触发继续 defer |
-| [Hierarchical Agent Stride Control v0](hierarchical-agent-stride-control-v0.zh-CN.md) | S11/S7 | 已接受；M1 只读观测 | P2：matched shadow stride 实验，定义代价与事件；不直接改变生产节奏 |
+| [Research Exploration Control Plane v0](research-exploration-control-plane-v0.zh-CN.md) | S11/S3 | 已接受；M2 composition/successor 局部实现；§11.5 result/adoption 修订为提案 | P1：一条真实 provider result -> 有范围 interpretation -> adopted step/successor，精确 lineage/shared write gate 与 packaged readback；shadow 资格先于 opt-in，保留 composition gap；推断与 M4 portfolio 排序继续 defer |
+| [Hierarchical Agent Stride Control v0](hierarchical-agent-stride-control-v0.zh-CN.md) | S11/S7 | 已接受；M1 只读观测；§7.5 checkpoint/monitor binding 为提案 | P2：matched shadow 决策时机、observer/model 成本与 lost-result recovery；保留每 Turn settlement，不改变默认 cadence |
 | [Goal-scoped Capability Portfolio v0](goal-scoped-capability-portfolio-v0.zh-CN.md) | S1/S3/S6/S8/S11 | 已接受；只读配置/上下文检查切片 | P1：复用既有 owner 验收纠正→新会话决策，再做按需组合与可测方法演化；不建万能记忆库或第二个启用开关 |
 | [Post-Outcome Memory Utility Attribution v0](post-outcome-memory-utility-attribution-v0.zh-CN.md) | S6/S11 | 已接受；Stage 1 verified-outcome 绑定 | P1 只读 reducer→P2 pilot：区分 recalled/applied/utility，归因不自动改 ranking |
 | [Obelisk Session Evidence Provider v0](obelisk-session-evidence-provider-v0.zh-CN.md) | S6/S8 | 已接受；可选只读评估 | P2：显式 gap recall、来源与权限可解析、关闭不影响主流程；不当 work authority |
diff --git a/docs/architecture/rfcs/research-exploration-control-plane-v0.md b/docs/architecture/rfcs/research-exploration-control-plane-v0.md
index 2c5bfa54ac..b7c32355fe 100644
--- a/docs/architecture/rfcs/research-exploration-control-plane-v0.md
+++ b/docs/architecture/rfcs/research-exploration-control-plane-v0.md
@@ -154,6 +154,7 @@ milestone status.
 | No exact obligation/Todo/result lineage | A research receipt is not proof of an authorized Todo transition or accepted Goal closure. |
 | Cold shadow not adopted by hot status/frontier | The existing #3173 projection remains behavior-compatible; canonical research obligations still need M3 integration. |
 | No dismissal or deferral contract | Evidence invalidation is visible, but typed candidate retirement and resumption remain unimplemented. |
+| No experiment-result adoption gate | Negative evidence can be recorded without proving that the next bound decision used its scope, validity or uncertainty. Proposed single-experiment integration is defined in §11.5; current cold receipts do not enforce it. |
 | Live qualification incomplete | Deterministic and real CLI/file-log tests establish state semantics, not model selection quality or scientific truth; no live Lark sync is qualified by projection tests. |
 | No promotion evidence for inferred combinations | Shared constraints are not known to be precise enough to trigger obligations. |
 
@@ -600,6 +601,119 @@ turn discovering which evidence command satisfies a protocol ceremony. A
 delivery receipt proves that context arrived; the observation proves whether
 the next action used it.
 
+### 11.5 Evidence-conditioned experiment decisions (proposed refinement)
+
+This 2026-10-04 proposal refines M3 before M4 portfolio selection. It does not
+claim a shipped experiment controller or change ordinary replan defaults.
+[Decision checkpoints](hierarchical-agent-stride-control-v0.md#75-long-running-work-and-decision-checkpoints-proposed-refinement)
+determine when a result needs attention; this section defines what a research
+decision must retain. Explore is the nearest capability owner. Native
+measurement, statistical interpretation and job/artifact handling stay with
+the domain or benchmark provider. Generic kernel rules do not interpret scores
+or infer research truth from Todo churn.
+
+#### Bind a question and the scope of its evidence
+
+For the first real caller, compose an explicit versioned Explore attachment;
+do not add unknown fields to research observation v0. Its conceptual data are:
+
+| Data | Owning meaning |
+|---|---|
+| Hypothesis and declared mechanism family | Stable Explore node/revision and explicit relation to a family node. Similar names/prose do not prove equivalence; a family is a scoped mechanism claim, not a kernel-inferred taxonomy. |
+| Probe and applicability | Planned intervention, candidate/input fingerprints, comparison, environment/data/evaluator revision, scope, resource/stop bounds. The provider declares what would support, contradict or leave the claim unresolved. |
+| Result and validity | Exact provider result revision/evidence, comparison and uncertainty/guardrail assessment. Failed, missing or incomparable measurements remain distinct from valid negative evidence. |
+| Decision and adoption | Exact input/result fingerprints, existing `continue`/`no_change`/`replan` path outcome, evidence refs and bound task step or grounded successor. A repeat needs a new discriminating question or an explicit replication/uncertainty reason and budget. |
+
+This is a design inventory, not an accepted wire schema or new global enum.
+Explore owns the future TS codec/reducer; compute/capture providers adapt it
+without a parallel decision owner. Reuse nodes, findings, `supports`/`refutes`,
+closure and input-fingerprint semantics. Compose a versioned decision
+attachment with the existing Goal path delta and Todo-bound recommendation
+receipts, rather than add another Next Action or settlement store. An ML
+adapter can map its existing hypothesis ledger's weakened/retired states;
+that vocabulary does not become a generic progress result class.
+
+| Evidence | Legal decision basis |
+|---|---|
+| Build or measurement failed | Repair the prerequisite or record the blocker; do not assert scientific refutation. |
+| One valid candidate is worse | Preserve the incumbent and bind evidence to this candidate/probe scope. Continue another probe, replicate or replan with reasons; do not automatically retire the family. |
+| Effect is within uncertainty or measurements conflict | Preserve unresolved/contradictory evidence; choose a bounded discriminating probe or justified replication. A small score decline is not drift. |
+| Declared coverage contradicts a hypothesis | Close/revise that exact claim with its closure basis. A successor cites the exclusion and what changes; a same-mechanism probe addressing an uncovered condition remains legal. |
+| Coverage supports family retirement | Retire only the declared family/scope with attributable evidence. New input, evaluator or applicability can reopen investigation; retirement never settles Goal acceptance by itself. |
+
+Rejecting one parameter setting can leave a family open; demonstrating failure
+to meet a declared constraint under covered conditions can justify a different
+family. The kernel validates attribution and transitions, not causal truth or
+experiment quality. Renaming identifiers or allocating a new evidence id for
+an unchanged measurement is not novelty. No fixed failure count forces a pivot.
+
+#### Make result adoption observable at the existing decision boundary
+
+1. Record/read back the validated result on the exact experiment node with
+   provider provenance and current input fingerprints. A launch receipt alone
+   is not research evidence.
+2. Under an explicit per-Goal opt-in, project one compact pending decision for
+   its bound experiment/Todo: question, incumbent, applicable exclusions,
+   uncertainty/invalidity, input/result refs and remaining budget. Building the
+   projection needs neither a full log nor a new reasoning-model call.
+3. The normal admitted Agent decides. Continuing is legal if it states what
+   new information the next probe should obtain and why the exclusion does not
+   already answer it. A pivot names the changed question/mechanism and its
+   evidence through the existing grounded successor path.
+4. Explore's typed reducer qualifies the decision attachment. The shared write
+   gate joins it to the exact Todo/replan/Turn and current input revisions,
+   committing it with the normal task-step or successor receipt. Generic
+   replan keeps its existing legal outcomes. A record/read, stale citation or
+   unrelated Todo cannot discharge this opt-in decision duty.
+5. Re-offer pending results after a crash; acknowledge only durable adoption.
+   Replay recovers existing decision receipts; same-Turn settlement remains
+   idempotent, while a newly admitted recovery Turn settles normally. Changed
+   inputs or results invalidate pending decisions; historical facts remain
+   available.
+
+The required relation is **result -> interpretation -> adopted action**.
+Binding is not proof of quality. Operational validity, new research information,
+native outcome improvement and Goal acceptance are separate observations. A
+useful negative result can be delivery without a score gain; a series of
+negative experiments must still expose changed knowledge and remaining frontier.
+
+#### Smallest usable path and qualification
+
+Start with the existing **benchmark-toolkit admitted-run monitor/closeout**
+seam, composed with an Explore experiment node. Its
+[`benchmark_runtime_observation_v0` and continuity reducers](../../../loopx/capabilities/benchmark_toolkit/README.md)
+already separate runtime liveness, attributable terminal results and launch
+generation. They classify provider facts; they neither run experiments nor
+establish result validity. A controlled provider must execute and recover one
+bounded build/evaluation trial, validate its result, deliver the pending
+decision, adopt an evidence-linked continuation/successor and read it back.
+Initially use a complete trial result; intermediate training checkpoints need
+their own provider postcondition and qualification. A preview-only domain
+pack is not an execution backend. This bounded M3
+prerequisite need not wait for inferred composition, a new multi-agent
+scheduler or M4 ranking, and does not close the remaining M3 composition work.
+
+First shadow-capture actual decisions against negative, inconclusive, failed
+and invalidated inputs without changing scheduling. Enforcement is a separate,
+reversible per-Goal opt-in in the existing Explore configuration owner.
+Default-off shared surfaces retain current behavior. Disabling stops new
+decision duties, preserves evidence and completes owed settlement; it does
+not delete job/result history or waive existing replan/authority obligations.
+
+The enforcing slice includes the existing frontend task/Explore journey:
+question, candidate/incumbent, waiting/result state, scoped negative evidence,
+proposed next action and reason; an authorized decision and its readback,
+including stale-result recovery. Optional Lark consumes the same projection
+and authority. A CLI-only prerequisite remains partial until the packaged
+interaction is qualified.
+
+Qualify real provider process and canonical state, revision/lease fences,
+duplicate/lost delivery, stop/revocation, missing and contradictory results,
+unjustified replay/renaming, over-retirement, changed inputs reopening a scope,
+valid same-family probes and default-off parity. Then independently qualify
+model adoption and scientific utility under matched total budgets. Lower
+control cost must not come from hiding useful negative evidence.
+
 ## 12. Write-Time Enforcement
 
 The write gate must reuse the same current goal-frontier and Explore gap
@@ -804,7 +918,7 @@ control-plane failures.
 | M0 | RFC, current-state inventory, and explicit ownership decision | Maintainer review; no runtime behavior | Accepted design |
 | M1 | Characterization fixtures plus typed research observation and closure contract in Explore | Deterministic normalization, privacy, compatibility, and negative tests | Implemented evidence/CLI slice; live research qualification remains separate |
 | M2 | Explicit-only composition candidate, canonical gap projection, and read-only status shadow | No pairwise inference; bounded packet; projection parity | Partial: #3173 legacy quota/successor; canonical binary cold shadow in CLI/Lark projection; hot status adoption and live Lark qualification remain |
-| M3 | Goal-frontier obligation, exact Todo/experiment lineage, and shared write-time gate | State/replay matrix and premerge canary pass | Not started |
+| M3 | Goal-frontier obligation, exact Todo/experiment lineage, shared write-time gate; proposed §11.5 single-experiment result/adoption prerequisite | State/replay matrix, real provider/state readback, default-off parity and packaged journey before enforcement | Not started; §11.5 is a design proposal, not delivered runtime |
 | M4 | Bounded multi-candidate cards, `composition_selection_v0`, real model-tool behavior qualification, and repeated live shadow | Model autonomously selects a legal semantic action from the delivered candidate set; selection quality is no worse than the declared fallback; compact receipts only | Not started |
 | M5 | Shared-constraint candidate ranking in shadow mode | Precision and cost evidence; no automatic trigger | Not started |
 | M6 | Optional inferred trigger | Explicit maintainer decision and measured promotion thresholds | Deferred |
@@ -829,7 +943,11 @@ composition-frontier/result-layer checks. It covers canonical reverse pairs,
 terminal coverage, attribution, replay, invalidated inputs, stale experiment
 lineage and real CLI/file-log readback. Projection tests do not establish live
 remote effects or model-selected research behavior. Delivery is tracked in
-[#5214](https://github.com/loopx-project/loopx/issues/5214); the issue remains open.
+[#5214](https://github.com/loopx-project/loopx/issues/5214), now closed. Its closure
+does not supply missing live/M3 qualification. Remaining research direction is
+tracked by [#4391](https://github.com/loopx-project/loopx/issues/4391) and
+[#3243](https://github.com/loopx-project/loopx/issues/3243); preserve the explicit
+milestone gaps here rather than create a parallel task tree.
 
 M3 is the first behavior-changing slice. It should be a separate PR so the
 obligation and write gate can be reviewed and reverted independently from the
@@ -918,6 +1036,7 @@ This is a living RFC, not an append-only diary.
 |---|---|
 | 2026-08-13 | Adopt Explore as the canonical research-topology owner; choose explicit-only composition candidates for v0; represent joint work as an experiment node; defer shared-constraint inference to shadow qualification. |
 | 2026-08-13 | Separate eligibility from ranking: the control plane owns a legal bounded candidate set, while the model autonomously prioritizes among multiple eligible candidates. A selection receipt proves a scheduling choice, not research truth. Defer the protocol to M4 rather than adding it to the #3173 runtime slice. |
+| 2026-10-04 | Propose §11.5 as a bounded M3 prerequisite: exact result -> scoped interpretation -> adopted task step/successor. Reuse Explore and task/replan receipts; keep Turn settlement, native scoring and scientific qualification independent. This RFC update delivers no automatic policy. |
 
 ## 20. Acceptance Criteria for the RFC
 
diff --git a/docs/architecture/rfcs/research-exploration-control-plane-v0.zh-CN.md b/docs/architecture/rfcs/research-exploration-control-plane-v0.zh-CN.md
index a497f608fa..56a2dc1b28 100644
--- a/docs/architecture/rfcs/research-exploration-control-plane-v0.zh-CN.md
+++ b/docs/architecture/rfcs/research-exploration-control-plane-v0.zh-CN.md
@@ -135,6 +135,7 @@ hypothesis”或“新 probe family”。它无法持久表达：A 和 B 都已
 | 没有精确 obligation/Todo/result lineage | Research receipt 不证明已授权 Todo transition 或已接受 Goal closure。 |
 | Cold shadow 未接入 hot status/frontier | 既有 #3173 投影保持行为兼容；canonical research obligation 仍需 M3 集成。 |
 | 没有 dismissal/deferral contract | Evidence invalidation 可见，但类型化 candidate retirement/resumption 尚未实现。 |
+| 没有 experiment-result adoption gate | 负证据可被记录，但不能证明下一绑定决策使用了其 scope、validity 与 uncertainty。§11.5 提议 single-experiment 集成；当前 cold receipt 不强制执行该提案。 |
 | Live qualification 不完整 | Deterministic 与真实 CLI/file-log 测试证明状态语义，不证明 model selection 质量或科学结论；projection 测试不构成 live Lark sync 资格。 |
 | inferred combination 没有 promotion evidence | 共享 constraint 的精度还不足以直接触发 obligation。 |
 
@@ -553,6 +554,101 @@ Host 应投影：
 满足协议仪式。Delivery receipt 证明 context 到达；observation 证明下一步是否
 真正使用了 context。
 
+### 11.5 由证据驱动的实验决策（修订提案）
+
+这项 2026-10-04 提案细化 M3，先于 M4 portfolio selection；不代表实验控制器
+已交付，不改变普通 replan 默认行为。
+[决策 checkpoint](hierarchical-agent-stride-control-v0.zh-CN.md#75-长作业与决策-checkpoint修订提案)
+决定何时关注结果，本节定义研究决策应保留什么。最近的 capability owner 是
+Explore；原生测量、统计解释和 job/artifact 处理留在 domain/benchmark provider。
+Generic kernel 不解释分数，也不从 Todo 更新推断研究真相。
+
+#### 绑定问题与证据适用范围
+
+第一个真实 caller 组合显式 versioned Explore attachment，不向 research
+observation v0 偷加 unknown field。概念数据如下：
+
+| 数据 | 所属语义 |
+|---|---|
+| Hypothesis 与已声明 mechanism family | 稳定 Explore node/revision，显式关联 family node。相似名字/prose 不证明等价；family 是有范围的机制假设，不是 kernel 推断的分类。 |
+| Probe 与 applicability | 计划 intervention、candidate/input fingerprint、比较对象、环境/数据/evaluator revision、scope、资源/停止边界。Provider 声明什么支持、反驳或不能确定该假设。 |
+| Result 与 validity | 精确 provider result revision/evidence、比较和 uncertainty/guardrail assessment。失败、缺失、不可比的测量与有效负证据区分。 |
+| Decision 与 adoption | 精确 input/result fingerprint，现有 `continue`/`no_change`/`replan` path outcome、evidence ref 与绑定 task step 或 grounded successor。重复 probe 要有新的判别问题，或明确的 replication/uncertainty 理由和预算。 |
+
+这是设计清单，不是已接受 wire schema 或新 global enum。未来 TS codec/reducer
+归 Explore；compute/capture provider 适配，不新增平行决策源。复用 node、finding、
+`supports`/`refutes`、closure 和 input-fingerprint 语义。Versioned decision
+attachment 与现有 Goal path delta、Todo-bound recommendation receipt 组合，
+不新增 Next Action 或 settlement store。ML adapter 可映射已有 hypothesis ledger
+的 weakened/retired 状态，但这套词表不成为 generic progress result class。
+
+| 证据 | 合法决策依据 |
+|---|---|
+| 构建或测量失败 | 修复前置条件或记录 blocker，不声称科学假设被反驳。 |
+| 一个有效 candidate 更差 | 保留 incumbent，证据绑定该 candidate/probe scope。带理由选择其他 probe、复验或 replan，不自动退休整个 family。 |
+| 效果处于 uncertainty 内或测量冲突 | 保留 unresolved/contradictory evidence，选择有界判别 probe 或有依据复验。小幅掉分不是 drift。 |
+| 已声明 coverage 反驳 hypothesis | 以 closure basis 关闭/修订精确假设；successor 引用排除结论与改变之处。针对未覆盖条件的同机制 probe 仍合法。 |
+| Coverage 支持退休 family | 只退休有 attributable evidence 的已声明 family/scope。新的 input、evaluator 或 applicability 可重开调查；退休不直接 settle Goal acceptance。 |
+
+否定一个参数仍可保留 family；在覆盖条件内证明不能满足已声明 constraint，
+才可能支持换 family。Kernel 验证 attribution 与 transition，不认证因果真相或
+实验质量。仅重命名 identifier 或为未变测量分配新 evidence id 不是 novelty。
+不用固定失败次数强制 pivot。
+
+#### 在现有决策边界让 adoption 可观察
+
+1. 在精确 experiment node 记录/读回 validated result，保留 provider provenance
+   与当前 input fingerprint。Launch receipt 单独不是 research evidence。
+2. 显式 per-Goal opt-in 时，为绑定 experiment/Todo 投影紧凑 pending decision：
+   问题、incumbent、适用 exclusion、uncertainty/invalidity、input/result ref 和
+   剩余预算。构造投影不需要完整日志或新增 reasoning-model 调用。
+3. 由正常获准 Agent 决策。继续合法，但须说明下一 probe 要获得什么新信息，
+   以及 exclusion 为什么尚未回答该问题。Pivot 指出改变的问题/机制与证据，
+   使用已有 grounded successor 路径。
+4. Explore typed reducer 验证 decision attachment。Shared write gate 关联
+   精确 Todo/replan/Turn 与当前 input revision，随普通 task-step/successor
+   receipt 提交。Generic replan 保留已有合法 outcome。Record/read、stale
+   citation 或无关 Todo 不能解除 opt-in decision duty。
+5. Crash 后重交 pending result，仅 durable adoption 后确认消费。Replay 恢复
+   原 decision receipt；同一 Turn 的结算重放幂等，新获准 recovery Turn 仍正常
+   结算。Input/result 改变使 pending decision 失效，历史事实继续保留。
+
+要求的是 **result -> interpretation -> adopted action**。绑定不证明质量。
+Operational validity、新研究信息、原生结果提升和 Goal acceptance 是不同
+observation。有效负实验可以是合法交付而没有涨分；连续负实验仍须暴露知识
+变化与剩余 frontier。
+
+#### 最小有用路径与资格验证
+
+先接现有 **benchmark-toolkit admitted-run monitor/closeout** seam，与 Explore
+experiment node 组合。其
+[`benchmark_runtime_observation_v0` 与 continuity reducer](../../../loopx/capabilities/benchmark_toolkit/README.md)
+已有 runtime liveness、可归属 terminal result 和 launch generation 的区分。
+它们分类 provider fact，不运行实验，也不认证 result validity。受控 provider
+须实际执行/恢复一次有界 build/evaluation trial、验证结果、交付 pending decision、
+adopt 带证据 continuation/successor 并读回。先使用完整 trial result；训练中间
+checkpoint 要另有 provider postcondition 与资格验证。Preview-only domain pack
+不是 execution backend。
+有界 M3 prerequisite 不必等待 inferred composition、新 multi-agent scheduler
+或 M4 排序，也不关闭 M3 剩余 composition 工作。
+
+先 shadow capture：按 negative、inconclusive、failed、invalidated input 评估
+实际决策，不改变 scheduling。Enforcement 是现有 Explore configuration owner
+中的独立、可回滚 per-Goal opt-in；默认关闭时共享 surface 保持当前行为。
+关闭停止派生新 decision duty，保留 evidence 并完成 owed settlement，不删除
+job/result history，不豁免已有 replan/authority obligation。
+
+Enforcing slice 包含现有 frontend task/Explore 旅程：问题、candidate/incumbent、
+waiting/result state、有范围负证据、拟议 next action 与理由；已授权决策和
+readback，包含 stale-result recovery。Optional Lark 消费同源 projection/authority。
+CLI-only prerequisite 在 packaged interaction 验证前仍是 partial。
+
+资格验证覆盖真实 provider process 与 canonical state、revision/lease fence、
+重复/丢失交付、stop/revocation、缺失/冲突结果、无依据 replay/改名、过度退休、
+input 改变后重开 scope、合法同 family probe 与 default-off parity。随后在
+matched 总预算下独立验证模型 adoption 和科学效用；降低控制成本不能靠隐藏
+有效负证据。
+
 ## 12. 写时强制
 
 Write gate 必须复用 quota 所用的同一 current goal-frontier 与 Explore gap
@@ -742,7 +838,7 @@ rule，以及 model variance 与 control-plane failure 的分离。
 | M0 | RFC、current-state inventory 与显式 ownership decision | Maintainer review；无 runtime behavior | 已接受的设计 |
 | M1 | Characterization fixture，以及 Explore 中的 typed research observation 与 closure contract | Deterministic normalization、privacy、compatibility 与 negative test | Evidence/CLI 切片已实现；真实研究 qualification 独立保留 |
 | M2 | Explicit-only composition candidate、canonical gap projection 与 read-only status shadow | 不做 pairwise inference；packet 有界；projection parity | 部分实现：#3173 legacy quota/successor；CLI/Lark projection 的 canonical binary cold shadow；hot status adoption 与 live Lark qualification 仍未完成 |
-| M3 | Goal-frontier obligation、精确 Todo/experiment lineage 与共享 write-time gate | State/replay matrix 与 premerge canary 通过 | 未开始 |
+| M3 | Goal-frontier obligation、精确 Todo/experiment lineage、共享 write-time gate；§11.5 提议 single-experiment result/adoption prerequisite | State/replay matrix、真实 provider/state readback、default-off parity；enforcement 前验证 packaged journey | 未开始；§11.5 是设计提案，不是已交付 runtime |
 | M4 | 有界 multi-candidate card、`composition_selection_v0`、真实 model-tool behavior qualification 与重复 live shadow | 模型从交付 candidate set 中自主选择合法 semantic action；选择质量不劣于 declared fallback；只保留 compact receipt | 未开始 |
 | M5 | Shared-constraint candidate 在 shadow mode 中排序 | 有 precision/cost evidence；不自动触发 | 未开始 |
 | M6 | 可选 inferred trigger | 显式 maintainer decision 与量化 promotion threshold | 延后 |
@@ -766,7 +862,11 @@ M1/M2 evidence 切片交付：
 result-layer 检查。覆盖反向配对 canonical identity、terminal coverage、attribution、
 replay、input invalidation、stale experiment lineage 和真实 CLI/file-log 读回。
 Projection 测试不证明 live remote effect 或模型自主研究行为。交付继续由
-[#5214](https://github.com/loopx-project/loopx/issues/5214) 跟踪，该 issue 保持打开。
+[#5214](https://github.com/loopx-project/loopx/issues/5214) 记录，该 issue 已关闭。
+关闭不补齐缺失的 live/M3 qualification。后续研究方向由
+[#4391](https://github.com/loopx-project/loopx/issues/4391) 与
+[#3243](https://github.com/loopx-project/loopx/issues/3243) 跟踪；本 RFC 保留明确
+milestone gap，不另建平行 task tree。
 
 M3 是第一个 behavior-changing slice。它应单独成 PR，使 obligation 与 write gate
 能够独立于 evidence schema 评审和回滚。
@@ -848,6 +948,7 @@ evidence-backed terminal result。
 |---|---|
 | 2026-08-13 | 采用 Explore 作为 canonical research-topology owner；v0 选择 explicit-only composition candidate；joint work 表示为 experiment node；shared-constraint inference 延后到 shadow qualification。 |
 | 2026-08-13 | 将 eligibility 与 ranking 分离：控制面拥有合法、有界 candidate set，模型在多个 eligible candidate 中自主择优；selection receipt 只证明调度选择，不构成 research truth。该协议延后到 M4，不进入 #3173 的首期 runtime。 |
+| 2026-10-04 | 提议 §11.5 作为有界 M3 prerequisite：精确 result -> 有范围 interpretation -> adopted task step/successor。复用 Explore 和 task/replan receipt；Turn settlement、原生评分与科学资格独立。本次 RFC 更新不交付自动策略。 |
 
 ## 20. RFC 验收标准