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 1500fc4351..74eaae5feb 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
@@ -1002,6 +1002,39 @@ delivery and model adoption qualify the path, not effectiveness. A dual-session
Controller/Worker runtime, numerical fidelity engine and learned policy are not
prerequisites. Later schema work needs an observed missing producer/consumer fact.
+
+
+#### Decision-episode readback for action-value evidence
+
+From a prospectively selected meaningful decision point, the first implementation
+slice should demonstrate one real, followable episode: **state then → pre-action
+forecast and choice → actual work/resources → returned evidence and next decision
+→ frozen terminal outcome**. Data semantics are in
+[research rationale §9](../../research/budgeted-research-planning.md#action-value-evidence);
+owners and negative cases are in
+[joint RFC §5.9 and C17–C18](target-yardstick-stride-decisions-v0.md#policy-evolution).
+Reuse existing plan/finding/measurement/followthrough. Report missing records as
+unrecoverable; do not fabricate retrospective forecasts or add model/Native calls
+to complete an episode. Include capture and analysis overhead in full cost.
+If the run is unfinished, leave the terminal link pending until the planned
+endpoint; do not add a run merely to complete readback.
+
+Readback should identify which key forecast was supported/refuted and how actual
+subsequent work addressed it. Evidence-backed continuation also counts as adoption.
+A two-hour local gain cannot label an eighteen-hour terminal forecast; an
+unobserved alternative is not zero. Keep the relevant work chain for delayed
+gains rather than crediting the last switch with all progress.
+
+This readback establishes linkage and analyzability. Synthetic examples can
+validate missing/late/revised-record handling but do not replace real adoption.
+Formal comparisons retain every prospectively selected episode, including
+failure, cancellation, no gain and insufficient evidence, rather than only
+success stories. Report forecast calibration, adoption/decision change,
+same-state paired contrasts, from-start terminal results and new-task benefit
+separately; passing one level does not pass the next. Hold out tasks, source runs
+or sibling-branch families for learning/calibration. Use the pilot and full-budget
+contracts below, not a separate experiment or expanded feedback permission.
+
#### Affordable conditional-effect pilot
Start with three complementary task types, under the pinned upstream task names:
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 5892fbdf1b..bc04714991 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
@@ -813,6 +813,28 @@ causal credit。给出强 utility 前必须做 holdout replay 或 bounded ablati
源码交付和模型采用只验证路径，不证明效果。双会话 Controller/Worker、数值 fidelity
引擎和学习型策略均非前置条件。后续 schema 修改要有已观察到的 producer/consumer 缺口。
+
+
+#### 动作价值证据的片段读回
+
+首个实现切片应从前瞻选定的有意义决策点，展示一个可接续的真实片段：
+**当时状态 → 事前预测与选择 → 实际工作／消耗 → 返回证据及下一决定 → 冻结终局**。
+数据语义见[研究说明 §9](../../research/budgeted-research-planning.zh-CN.md#action-value-evidence)，
+owner 与负例见[联合 RFC §5.9、C17–C18](target-yardstick-stride-decisions-v0.zh-CN.md#policy-evolution)。
+复用已有 plan/finding/measurement/followthrough；原记录缺失时报告不能恢复，
+不事后补造预测，不为补齐片段增加模型或 Native 调用。采集与分析开销进入完整成本。
+若全程尚未结束，终局先保留为待关联，在预定终点结算；不为完成读回追加一轮运行。
+
+读回需要明确关键预测有没有被支持或推翻，以及实际下一段工作如何处理它；
+有据继续原计划也是采用。两小时后的局部涨分不能充当十八小时终局标签，
+未选路线没有结果也不填零。延迟增益保留相关工作链，而非全部归给最后一次 switch。
+
+这条读回证明的是关联与可分析性；合成示例可验证缺失／迟到／修订处理，不能替代真实采用。
+正式比较保留全部预选片段，包括失败、取消、无收益和证据不足，避免只展示成功故事。
+分别报告预测校准、采用与决策变化、同状态配对差异、从头终局及新任务收益；
+前一层通过不自动使后一层通过。学习／校准留出隔离任务、来源 run 或同父分叉家族，
+仍沿用下述 pilot 与完整预算合同，不另起一套实验或放宽反馈权限。
+
#### 可负担的条件效应 pilot
先选三个互补任务类型，使用固定 upstream 版本的任务名称：
diff --git a/docs/architecture/rfcs/target-yardstick-stride-decisions-v0.md b/docs/architecture/rfcs/target-yardstick-stride-decisions-v0.md
index afd1daee70..09cbc5e956 100644
--- a/docs/architecture/rfcs/target-yardstick-stride-decisions-v0.md
+++ b/docs/architecture/rfcs/target-yardstick-stride-decisions-v0.md
@@ -567,6 +567,8 @@ population-selection operators remain `advisory_only`
(`execution_authority_granted=false`) until matched-budget evidence qualifies
promotion.
+
+
### 5.9 Policy evolution through the same owner
| Stage | Useful next result | Retained boundary |
@@ -583,6 +585,33 @@ explicitly justified off-policy assumptions for policy-value claims. Preserve
failed, cancelled and unknown results. Raw task data stays within the research
boundary and cannot become a production memory or training corpus by default.
+When consuming these records for later learning, use the
+[action-value evidence model](../../research/budgeted-research-planning.md#action-value-evidence)
+to separate observed progress, value predicted under a continuation and remaining
+budget, and the causal contrast against a feasible alternative. A gain after
+`switch` directly supports only the first. An action name does not replace a
+concrete method, and an unchosen route is not a zero-score label.
+
+For selected decision episodes, existing provenance/followthrough should recover:
+
+- State/evidence visible at selection and remaining budget, compared routes,
+ concrete selected method, continuation revision and forecast horizon.
+- Key predictions, their basis or unknowns, and observations supporting continuation
+ or revision, recorded before outcomes are visible; keep original forecasts on revision.
+- Actual method, revision, resources and deviations; the observation's original
+ candidate/evaluator, availability time, and resulting next decision.
+- Related prerequisite and subsequent work, and the terminal outcome under the
+ frozen horizon/selection rule; preserve failure, cancellation, pending and missing coverage.
+
+These are proposed training-example semantics, not new mandatory fields on every
+Turn. Mark facts unrecoverable from existing references as missing; do not backfill
+them from later summaries or trigger evaluations to complete a form. Test forecast
+calibration, actual adoption, local paired contrasts and frozen-policy new-task
+benefit separately. Hold out shared tasks/source runs or whole branch families,
+not randomly mixed adjacent Turns. Continuing a route supported by evidence also
+counts as adoption; cue delivery, complete labels or a direction change alone
+do not establish policy benefit.
+
No new policy registry, plan database, scheduler or universal scalar rank is
needed for the first caller. Before any later schema change, demonstrate the
missing producer/consumer fact, extend the existing versioned owner, and qualify
@@ -675,6 +704,8 @@ These are implementation acceptance requirements, not tests already passed.
| C14 | Same evidence, different remaining horizons | Compare feasible terminal outcomes, including confirmation; do not require the same route or use a fixed per-candidate score. |
| C15 | A prerequisite has no immediate gain but enables a plausible later improvement | Consider the conditional sequence and a bounded discriminating first slice; count dependent milestones once as state evidence, not independent rewards. |
| C16 | Confirmation is pending or cannot finish under the terminal protocol | Retain immutable result attribution and honest unknowns; consider independent work or a feasible validated route without resetting cooldown or inventing gain. |
+| C17 | The next measurement improves, but the alternative is unobserved and earlier work also contributed | Record observed progress only; qualify predicted action value by continuation/budget, leave the causal contrast unknown, and link prerequisites instead of labeling the last action with all gain. |
+| C18 | A forecast is revised after the outcome, feedback arrives late, or related episodes cross training/evaluation | Recover forecasts/evidence from before result visibility; revisions preserve originals, local measurements do not stand for terminal outcomes, and evaluation isolates tasks/source runs or branch families. |
Replay typed scenarios with synthetic evidence against real File/SQLite owners
where persistence is touched. Use actual host/CLI settlement and packaged
diff --git a/docs/architecture/rfcs/target-yardstick-stride-decisions-v0.zh-CN.md b/docs/architecture/rfcs/target-yardstick-stride-decisions-v0.zh-CN.md
index a0c19ca560..cf60116e78 100644
--- a/docs/architecture/rfcs/target-yardstick-stride-decisions-v0.zh-CN.md
+++ b/docs/architecture/rfcs/target-yardstick-stride-decisions-v0.zh-CN.md
@@ -418,6 +418,8 @@ Bellman 表达式不会提供状态转移模型；事后峰值不能成为决策
而非按特性复制校验或排序样板（配套收敛提案见 [#8516](https://github.com/loopx-project/loopx/pull/8516)）。
分析与种群选择算子在通过匹配完整预算验证前一律保持 `advisory_only`（`execution_authority_granted=false`）。
+
+
### 5.9 沿同一 owner 演进策略
| 阶段 | 下一项有用结果 | 保留的边界 |
@@ -431,6 +433,22 @@ Bellman 表达式不会提供状态转移模型；事后峰值不能成为决策
策略价值主张使用前瞻匹配接续，或明确论证离线策略评估的假设。保留失败、取消和未知结果。
原始任务数据留在研究边界内，不因记录轨迹就进入生产 memory 或训练语料。
+为后续学习消费这些记录时，按[动作价值证据](../../research/budgeted-research-planning.zh-CN.md#action-value-evidence)
+区分三件事：实际路线上的进展、给定后续规则与剩余预算的价值预测、相对可行替代的因果差异。
+`switch` 后涨分只能直接支持第一项；动作名不能代替具体方法，未选路线不能补成零分。
+
+在选定的决策片段中，已有 provenance/followthrough 应支持恢复：
+
+- 选择时可见的状态／证据与剩余预算，以及所比较路线、具体所选方法、后续策略版本和预测期限。
+- 结果可见前记录的关键预测、依据或未知项，以及会支持继续或修订的观察；后续修订保留原预测。
+- 实际方法、版本、资源与计划的差异；返回观察的原候选／评测归属、可见时刻及由此产生的下一决定。
+- 相关前置和接续工作，以及冻结期限／选择规则下的终局；失败、取消、pending 与覆盖缺失分别保留。
+
+这些是拟议学习样本的语义，不为所有 Turn 新增必填项。现有引用不能恢复的事实标为缺失，
+不要从后来摘要倒填，或为满足表格额外触发评测。分开检验预测校准、实际采用、局部配对差异与
+冻结策略的新任务收益；留出划分隔离共享任务／来源运行或分叉家族，不随机拆散相邻 Turn。
+有证据支持原路线时，继续也是采用；提示送达、标签完整或方向改变都不单独证明策略有效。
+
首个调用方不需要新策略 registry、计划库、scheduler 或通用标量排序。
以后若修改 schema，先证明缺失的生产者/消费者事实，再扩展既有版本化 owner，
验证读回、恢复和关闭状态。Controller–Worker 组合以后可消费同一决策，
@@ -508,6 +526,8 @@ Provider 只暴露授权证据。原始任务、trace、凭据、私有 evaluato
| C14 | 同样证据、不同剩余时长 | 比较包含确认的可行终局结果，不要求选择同一路线或使用固定的候选自身分数。 |
| C15 | 前置工作不立即涨分，但使后续改善成为可能 | 考虑条件序列与有界的首个判别切片；相互依赖的里程碑只作为状态证据，不当成独立奖励。 |
| C16 | 确认仍在进行，或无法在终局协议内完成 | 保留不可变结果归属和真实 unknown；比较独立工作或可行的已验证路线，不重置 cooldown 或虚构收益。 |
+| C17 | 下一次测量涨分，但无替代路线结果，且更早工作共同贡献 | 只记录所观测进展；动作价值注明后续策略／预算，因果差异保持未知，关联前置工作，不把全部增量当作最后动作标签。 |
+| C18 | 预测被事后修订、反馈迟到，或同源片段跨训练与评测 | 恢复结果可见前的预测和证据；修订不覆盖原记录，局部测量不冒充终局，评估按任务／来源运行或分叉家族隔离。 |
若涉及持久化，用合成证据在真实 File/SQLite owner 上重放 typed 场景。
首次采用要覆盖实际 host/CLI settlement 和 packaged readback；schema 单测不足以验收。
diff --git a/docs/research/budgeted-research-planning.md b/docs/research/budgeted-research-planning.md
index 2fd7d72af8..b68eccd9b8 100644
--- a/docs/research/budgeted-research-planning.md
+++ b/docs/research/budgeted-research-planning.md
@@ -259,6 +259,15 @@ frozen terminal rule. Do not manufacture extra evaluation opportunities.
## 6. Why action logs are not causal labels
+"At state s, switch was selected; two hours later the next Native result improved
+by one point" is useful evidence, but not "switch added one point over continuing
+grow." If a matched grow branch improved by two points, switch made +1 progress
+but had a −1 contrast. Without executing or credibly identifying the comparator,
+that contrast remains unknown. Two hours may only be the interval to the next
+measurement, not the remaining research horizon. Earlier implementation,
+calibration and exploration may also contribute; the last action does not own
+the whole observed gain.
+
The observed fraction of gains after grow, verify or switch mixes different
starting states. Grow may be selected after promising evidence; switch after a
route stalls. Their raw frequencies cannot rank the actions at one shared state.
@@ -319,6 +328,13 @@ and cost → observation → revised continuation → accepted terminal outcome*
Keep failed/cancelled work, delayed results and unknowns. Attribute dependencies
across the sequence instead of giving all terminal gain to its last action.
+The most useful feedback explains **why the action seemed worthwhile then,
+which key forecast was supported or refuted, and how that changed the next
+choice**. A tail counterexample can refute the premise for a wider tuning sweep;
+subsequent work should address that question. Evidence-backed continuation also
+counts as adoption when new evidence supports the original route; switching is
+not mandatory.
+
## 8. Existing owners and the smallest useful next experiment
@@ -362,9 +378,83 @@ supported by trajectories showing what decision changed and why. More complete
forms, extra Native calls or a plausibly intelligent explanation are insufficient.
Detailed task choice and run authorization remain in the pilot, not this essay.
+
+
## 9. From useful feedback to a learned research policy
-Use the existing decision/evidence owners through three stages:
+### 9.1 Learning target: the current action and remaining continuation
+
+A Meta Model used to allocate research budget should estimate value conditional
+on state, remaining budget and the subsequent policy, not assign grow/verify/switch
+one fixed success rate each. Let Y_b be the terminal outcome selected under the
+frozen rule within budget b, a the concrete first action, and π the continuation:
+
+$$
+Q^\pi(s,b,a)=\mathbb E[Y_b^{a,\pi}\mid S_0=s],\qquad
+A^\pi(s,b;a,a_0)=Q^\pi(s,b,a)-Q^\pi(s,b,a_0).
+$$
+
+The first quantity forecasts the terminal result after starting this way and
+continuing with π; the second compares another feasible first route. Both include
+the first action's resource use. Specify a by method, candidate and verification
+choice; a coarse action name does not define a treatment. A changed π revision,
+remaining budget or acceptance rule changes the prediction target. One completed
+trajectory supplies a noisy outcome for the executed route, not the true Q or a
+complete label for A.
+
+Start with checkable subpredictions: what a probe will show, implementation
+duration, or whether confirmation can meet the deadline. Then test whether they
+help choices. Better forecasts need not change action ranking, and better local
+ranking need not improve the terminal result. Stride chooses how far to pursue
+a direction, Yardstick identifies measurement scope and confirmation gaps, and
+the Meta Model estimates conditional value for those choices. Estimates do not
+replace the existing owners' authority, budget or acceptance checks.
+
+### 9.2 A reconstructable decision episode
+
+Link the following at meaningful decision boundaries; do not require a form per
+tool call or a full-trajectory reread every Turn. Reconstruct from existing
+decision, finding, execution and measurement references. These are semantic
+requirements, not a new wire schema.
+
+| Time | What must be recoverable | Why it matters |
+| --- | --- | --- |
+| Before selection | Visible candidates/incumbent, directions, premises/counterevidence, verification gaps, pending results and remaining resources; source revisions and availability times | Forecast from information available then, not a later rewritten summary |
+| Forecast and choice | A few feasible alternatives, selected concrete method, key result and duration predictions/ranges, basis, and what would change the next choice | Separate the prior reason, uncertainty and retrospective explanation; do not invent precise probabilities |
+| Execution | Actual method versus plan and reason for changes, artifact revision, complete resources, failure/cancellation and waiting | Proposal is not adoption, adoption is not success; do not double-count overlapping wall time |
+| Feedback and next decision | Candidate/workload/evaluator of the observation, when feedback became available, forecast supported/refuted, and actual continuation/revision | Late results belong to the originating experiment; a decision uses only evidence then available; pending is not zero |
+| After the episode | Continuation revision, relevant prerequisite/subsequent work, terminal result at the frozen horizon and selection rule | Preserve delayed value and multi-step dependencies instead of crediting only the last action |
+
+Record a source-backed forecast revision before the result is visible. Retain
+old predictions and new evidence when revising, rather than overwriting the
+forecast. A two-hour measurement checks the corresponding horizon; an eighteen-hour
+terminal forecast needs that endpoint or the predeclared missing/cancelled-outcome
+treatment. Do not relabel a local gain or pending result as terminal success/failure.
+
+### 9.3 Four data sources with different evidential roles
+
+| Data | What it can test | What it does not establish |
+| --- | --- | --- |
+| Pre-action forecasts and corresponding realized outcomes | Probability, gain-range and duration calibration on covered states; report missing/cancelled outcomes separately | Unchosen outcomes, or that accurate prediction makes allocation better |
+| Same-candidate Proxy–Native pairs | Measurement bias, ordering and scope | Proxy improvement as a causal policy gain |
+| Different continuations from the same checkpoint | Local contrasts and uncertainty under equivalent restoration, isolation and aligned budgets | One pair as an exact expectation, or the policy's from-start effect |
+| Complete new-task runs with a frozen policy | Matched-budget terminal benefit and transfer | Repeated same-task tuning as independent generalization evidence |
+
+Specify the forecast target, horizon, scope and missing-outcome treatment before
+choosing calibration metrics. Retain all prospectively selected episodes, not
+only successful forecasts. Calibration data follows the actions actually chosen
+and usually does not cover every alternative. Off-policy evaluation separately
+needs coverage and comparability assumptions; unchosen actions are not zero-reward
+labels. Adjacent episodes from one task and branches sharing a prefix are
+dependent: hold out tasks/source runs or whole branch families rather than
+randomly splitting Turns. Shared parent trajectories can cross the training/evaluation
+split; future peaks, later test answers and post-hoc explanations must not enter
+decision-time features either.
+
+### 9.4 Validate progressively through existing owners
+
+First test whether the current model uses this feedback chain to improve
+EdgeBench continuation; then decide whether a separate learned model is useful:
| Stage | Learn or demonstrate | Evidence that can support it |
| --- | --- | --- |
@@ -372,18 +462,11 @@ Use the existing decision/evidence owners through three stages:
| Calibrated prediction | More reliable gain, duration and failure forecasts | Predictions recorded before outcomes, evaluated on held-out episodes |
| Learned allocation | Better policies for state and remaining budget | Prospective policy comparisons and frozen new-task tests |
-Each training example should connect what was visible at decision time,
-considered alternatives and predictions, the selected action and actual work,
-resource use, successor evidence and later terminal outcome. These are semantic
-requirements on useful data, not additional unsupported fields on current v0.
-Keep task/source permissions and candidate/version lineage.
-
-Selected-action logs can train predictors but do not reveal every unchosen
-outcome. Proxy/Native pairs can calibrate measurement relationships, and paired
-forks can estimate local treatment differences; neither automatically supplies
-all action values. Offline policy evaluation needs its own coverage and
-identification assumptions. Future peaks, later test answers and post-hoc
-interpretations must not leak into decision-time features.
+The pilot's [episode readback and separate acceptance levels](../architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.md#action-value-episode)
+use the existing experimental budget and feedback protocol. Preserve task/source
+permissions and candidate/version lineage; diagnostic records do not authorize
+training. A more complete model or full trajectory warehouse is not a prerequisite
+for the first useful comparison.
For an implementation or document handoff, preserve the chain:
**objective and comparator → theoretical reason → existing owner → proposed
diff --git a/docs/research/budgeted-research-planning.zh-CN.md b/docs/research/budgeted-research-planning.zh-CN.md
index 1861194eb8..7b081738cd 100644
--- a/docs/research/budgeted-research-planning.zh-CN.md
+++ b/docs/research/budgeted-research-planning.zh-CN.md
@@ -206,6 +206,12 @@ Native 确认可用于确认候选、改变研究方向或校准代理，其价
## 6. 为什么动作日志不是因果标签
+“在状态 s 选择 switch，用了两小时，下一次 Native 提高 1 分”是一条有用记录，
+但不是“switch 比当时继续 grow 多带来 1 分”。假设同起点、同预算的 grow 分支提高
+2 分，则 switch 自身进展为 +1，相对 grow 却是 −1；没有执行或可信识别对照时，
+这个差值仍是未知。两小时还可能只是决策到下一次测量的间隔，不是整个剩余研究期限。
+下一次提分也可能依赖更早的实现、校准和探索；不能把它全部归给最后一步。
+
grow、verify、switch 后的涨分比例，混合了不同起点。
grow 可能发生在积极证据之后，switch 可能发生在路线停滞之后。
原始频率不能给同一状态下的动作排序；incumbent 不变，也不能证明某次检查防住了回归。
@@ -257,6 +263,10 @@ checkpoint 分叉估计的是该状态下的继续效果，不是完整策略从
修订接续 → 验收终局成果**。保留失败、取消、迟到结果和未知，关联跨步骤依赖，
不要把终点全部增益归给最后一个动作。
+这里最有价值的反馈是：**当时为什么值得做、哪个关键预测被支持或推翻，以及它怎样
+改变后续选择**。例如尾部反例推翻了“可直接扩大调参”的前提，下一段工作应能看见对此
+问题的处理；新证据支持原路线时，有据继续也算反馈被采用，不强求切换。
+
## 8. 已有 owner 与最小有用实验
@@ -291,9 +301,67 @@ checkpoint 分叉估计的是该状态下的继续效果，不是完整策略从
更完整的表单、更多 Native 调用、听起来聪明的理由都不足以证明。
详细选题与运行授权留在 pilot 合同，不由本文另行决定。
+
+
## 9. 从有用反馈到可学习的研究策略
-沿同一决策／证据 owner 演进：
+### 9.1 学习目标：当前动作和剩余接续的价值
+
+Meta Model 若用于分配研究预算，应估计给定状态、剩余预算和后续策略时的价值，
+而非把 grow／verify／switch 当成各有一个固定成功率的类别。令 Y_b 为预算 b 内按
+冻结规则选出的终局成果，a 为当前具体行动，π 为此后的接续规则：
+
+$$
+Q^\pi(s,b,a)=\mathbb E[Y_b^{a,\pi}\mid S_0=s],\qquad
+A^\pi(s,b;a,a_0)=Q^\pi(s,b,a)-Q^\pi(s,b,a_0).
+$$
+
+第一项预测“这样开始、随后按 π 继续”的终局；第二项比较当前另一条可行路线。
+它们都包含当前动作的资源消耗。a 应明确到具体方法、候选与验证方式，粗粒度动作名
+不够定义处理。π 的版本、剩余预算或验收规则改变，预测目标也随之改变。
+单条完成轨迹只提供所执行路线的一个带噪声结果，不是真实 Q，更不是 A 的完整标签。
+
+可以先学可核验的子预测：探针会出现什么结果、实现要多久、确认能否赶上期限；
+再检验它们是否帮助选择。更准确的预测未必改变排序，更合理的局部排序也未必提升终局。
+Stride 决定一次沿方向推进多深，Yardstick 判断测量可信范围与确认缺口，Meta Model
+为这些选择提供条件价值估计；估计不替代现有 owner 的权限、预算与验收。
+
+### 9.2 一个可恢复的决策片段
+
+在有意义的决策边界关联以下内容，不要求每个工具调用都填表，也不要求每轮重读完整轨迹。
+用现有 decision、finding、执行与测量引用重建；下表是语义要求，不是新增 wire schema。
+
+| 时点 | 要能恢复的内容 | 为什么需要 |
+| --- | --- | --- |
+| 选择前 | 当时可见的候选／incumbent、方向、前提与反证、验证缺口、待返回结果、剩余资源；来源版本与可见时刻 | 保证预测使用的是当时已有信息，而非后来重写的摘要 |
+| 预测与选择 | 少量可行替代、所选具体方法、关键结果与耗时的预测／范围、依据、什么结果会改变下一步 | 区分事前理由、模型不确定性与事后解释；不凭空填精确概率 |
+| 实际执行 | 实际方法与计划的差异及原因、产物版本、完整资源消耗、失败／取消与等待 | 提议不是采用，采用不是成功；并行墙钟不重复累计 |
+| 反馈与下一决定 | 观测的候选／workload／evaluator、反馈何时可用、哪项预测被支持或推翻、实际继续／修订了什么 | 迟到结果归原实验，新决策只能使用其当时可见的证据；pending 不填零 |
+| 片段之后 | 后续规则版本、相关前置与接续工作、冻结期限和选择规则下的终局结果 | 保留延迟收益与多步依赖，避免把所有增益归给最后动作 |
+
+预测应在结果可见前留下有来源的版本；修订保留旧值和新证据，而不是覆盖原预测。
+一次两小时后的测量只能检验相应时间窗的预测；十八小时终局预测要等其终局或按
+预先声明的缺失／取消规则处理，不能把局部上涨或未返回结果改标为终局成功／失败。
+
+### 9.3 四类数据，四种证据能力
+
+| 数据 | 可检验的内容 | 不应推出的结论 |
+| --- | --- | --- |
+| 事前预测及对应实际结果 | 在已覆盖状态上校准概率、收益范围与耗时；分别报告缺失和取消 | 已经知道未选动作的表现，或预测准确就证明分配更优 |
+| 同候选的 Proxy–Native 配对 | 校准测量偏差、排序与适用范围 | proxy 改善就是策略因果收益 |
+| 相同 checkpoint 的不同后续路线 | 在恢复等价、隔离和预算对齐下估计局部差异及不确定性 | 一对结果等于精确期望，或已证明从零运行的总效应 |
+| 冻结策略后的新任务完整运行 | 检验等预算终局收益与迁移 | 同题反复调参后的优势是独立泛化证据 |
+
+先明确预测对象、时间窗、适用范围和缺失结果处理，再选择校准指标；保留全部预选片段，
+不能只挑成功预测。校准数据来自实际被选择的行动，通常不覆盖所有候选路线。
+离线策略评估另需行为覆盖与可比性假设，不把未选行动填成零奖励。
+同一任务的相邻片段、共享前缀的分叉高度相关；留出划分应按任务／来源运行或分叉家族
+隔离，而不是随机打散 Turn。共享父轨迹可能跨越训练／评估分区；未来峰值、后来的
+测试答案和事后解释也不能进入决策当时的特征。
+
+### 9.4 沿现有 owner 逐步验证
+
+近期先验证当前模型能否利用这条反馈链改善 EdgeBench 接续，再决定是否训练独立模型：
| 阶段 | 学习或证明什么 | 能支持它的证据 |
| --- | --- | --- |
@@ -301,14 +369,9 @@ checkpoint 分叉估计的是该状态下的继续效果，不是完整策略从
| 校准预测 | 更可信的增益、耗时和失败预测 | 结果之前记录的预测，在留出 episode 上评估 |
| 学习分配 | 针对状态与剩余预算的更好策略 | 前瞻策略比较、冻结后的新任务测试 |
-每条学习样本应连起决策时可见的内容、考虑过的替代与预测、所选动作和实际工作、
-资源消耗、后继证据及后续终局成果。这些是有用数据的语义要求，不是给当前 v0
-加入不支持的字段。保留任务／来源权限，以及候选／版本关联。
-
-已选动作日志可以训练预测器，却不会揭示每个未选动作的结果。
-Proxy–Native 配对可校准测量关系，配对分叉可估计局部处理差异；二者都不自动提供所有动作价值。
-离线策略评估需要自己的覆盖与识别假设。未来峰值、后来的测试答案、事后解释，
-不能泄漏进决策当时的特征。
+首个 pilot 的具体[片段读回与分层验收](../architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.zh-CN.md#action-value-episode)
+沿用现有实验预算与反馈协议。保留任务／来源权限及候选／版本关联；记录可用于诊断，
+不自动授权训练。更完整的模型或全量轨迹库不是第一轮有用比较的前置条件。
接续实现或改写方案文档时，保留这条论证：
**目标与对照 → 理论理由 → 现有 owner → 拟议改变 → 什么证据能证明收益**。