huangruiteng · Oct 18, 2026 · Oct 18, 2026 · Oct 18, 2026
@@ -150,7 +150,7 @@ export function useResearchAnalysis(goalId: string, active: boolean, zh: boolean
   };
   const closeCitation = () => {citeGeneration.current++; setCitation(null); setCitationError(""); setCitationBusy(false); requestAnimationFrame(() => sourceTrigger.current?.focus());};
   return {readOnly: Boolean(readSourceUrl), citation, citationBusy, citationError, sourcePanel, readCitation, closeCitation, goalId, zh, location, snapshot, pack, readError, error: error || readError, loading, pending, cancelPending, recordedRunning, panel, trigger, read, show, execute, cancel,
-    close: () => {citeGeneration.current++; setCitation(null); setCitationError(""); setCitationBusy(false); navigate(false); requestAnimationFrame(() => trigger.current?.focus());}, select: (id: string) => navigate(location.open, id)};
+    close: () => {citeGeneration.current++; setCitation(null); setCitationError(""); setCitationBusy(false); navigate(false); requestAnimationFrame(() => trigger.current?.focus());}, select: (id: string) => navigate(location.open, id, true)};
 }
 type Controller = ReturnType<typeof useResearchAnalysis>;
 

@@ -8,6 +8,20 @@
 .research-progress select { min-width: 0; max-width: 400px; min-height: 44px; border: 1px solid var(--pw-line); border-radius: 6px; padding: 8px; background: var(--pw-bg); color: var(--pw-text); font: inherit; }
 .research-progress p, .research-progress small { color: var(--pw-muted); font-size: var(--text-caption); line-height: var(--leading-prose); }
 .research-progress summary { min-height: 44px; padding-block: 12px; cursor: pointer; font-size: var(--text-caption); }
+.research-retained-improvement { margin-block-start: 20px; }
+.research-retained-improvement h4 { margin: 0; font-size: var(--text-body); font-weight: 500; }
+.research-retained-improvement summary { display: flex; align-items: baseline; flex-wrap: wrap; gap: 8px 12px; list-style: none; }
+.research-retained-improvement summary::-webkit-details-marker { display: none; }
+.research-retained-improvement summary strong { font-size: var(--text-title); font-weight: 600; font-variant-numeric: tabular-nums; }
+.research-retained-improvement summary span { color: var(--pw-muted); font-size: var(--text-body); }
+.research-retained-improvement details[open] summary svg { transform: rotate(180deg); }
+.research-retained-improvement dl { margin: 0; font-size: var(--text-body); }
+.research-retained-improvement dl > div { padding-block: 6px; }
+.research-retained-improvement dt { color: var(--pw-muted); font-size: var(--text-caption); }
+.research-retained-improvement dd { margin: 0; }
+.research-retained-improvement dd button { display: inline-flex; align-items: center; gap: 8px; min-height: 44px; padding: 0; background: none; border: 0; color: var(--pw-text); font: inherit; cursor: pointer; }
+.research-retained-improvement dd button:hover { text-decoration: underline; }
+.research-retained-improvement summary:focus-visible, .research-retained-improvement dd button:focus-visible { outline: 2px solid var(--pw-blue); outline-offset: 2px; }
 .research-effect-metrics { display: grid; grid-template-columns: repeat(6, minmax(0, 1fr)); gap: 8px 24px; margin-block-start: 12px; }
 .research-effect-metric { min-width: 0; }
 .research-effect-metric[open] { grid-column: 1 / -1; }

@@ -30,6 +30,28 @@ export function verdictLabel(metric: TrajectoryMetric, verdict: AssessmentVerdic
 
 // Format the native owner's existing projection, without deriving a gain,
 // deciding adoption, or borrowing values from another pack.
+function RetainedImprovement({native, zh, onCitation}: {native: NonNullable<ResearchProgressPack["native_outcomes"]>; zh: boolean; onCitation: (citation: AssessmentCitation, origin?: HTMLButtonElement) => void}) {
+  const outcomes = native.projection, improvement = outcomes.retained_improvement;
+  if (!improvement) return null;
+  const format = (value: number) => new Intl.NumberFormat(zh ? "zh-CN" : "en", {maximumSignificantDigits: 10, signDisplay: "exceptZero"}).format(value);
+  const label = zh ? "已采用成果增量" : "Retained improvement";
+  return <section className="research-retained-improvement" aria-label={label}>
+    <div className="research-progress-title"><h4>{label}</h4><ContextHelp label={zh ? "成果增量说明" : "About retained improvement"}>
+      <p>{zh ? "固定基线与实际保留结果的原生指标差值，正数表示改善。越低越好的指标反向计算；回退到基线为零，缺少基线或保留证据为未知。" : "Native change from a fixed baseline to the actually retained result. Positive means improvement; lower-is-better metrics reverse the subtraction. Returning to baseline is zero; missing baseline or retention evidence is unknown."}</p>
+      <p>{zh ? "只代表所选比较，不是 LoopX 的独立贡献或原目标已达成；不能跨题、跨单位或重叠窗口相加。展开数值可看原文。" : "This comparison alone does not establish LoopX's causal contribution or objective completion. Do not sum across tasks, units or overlapping windows. Expand the value for original evidence."}</p>
+    </ContextHelp></div>
+    {native.applicability === "historical" ? <p className="research-progress-warning" role="status">{zh ? "历史捕获 · 当前结果未核验" : "Historical capture · current result unverified"}</p> : null}
+    <details><summary aria-label={`${label}: ${improvement.value === null ? (zh ? "未知" : "Unknown") : `${format(improvement.value)} ${outcomes.metric.unit}`}`}>
+      <strong data-retained-improvement>{improvement.value === null ? "—" : format(improvement.value)}</strong>
+      <span>{improvement.value === null ? (zh ? "未知" : "Unknown") : outcomes.metric.unit} · {outcomes.metric.name}</span><ChevronDown size={14} aria-hidden/>
+    </summary><dl>
+      {([[improvement.baseline, zh ? "比较基线" : "Comparison baseline"], [outcomes.retained, zh ? "实际保留" : "Actually retained"]] as const).map(([point, title]) => <div key={title}><dt>{title}</dt><dd>{point ? <button type="button" onClick={event => onCitation(point.citation, event.currentTarget)}>{String(point.value)} · {point.candidate_ref}<ArrowUpRight size={13} aria-hidden/></button> : (zh ? "未知" : "Unknown")}</dd></div>)}
+      <div><dt>{zh ? "保留依据" : "Retention evidence"}</dt><dd>{improvement.retention_citation ? <button type="button" onClick={event => onCitation(improvement.retention_citation!, event.currentTarget)}>{zh ? "查看原文" : "Read original evidence"}<ArrowUpRight size={13} aria-hidden/></button> : (zh ? "未采集" : "Uncaptured")}</dd></div>
+      <div><dt>{zh ? "最后候选差值" : "Final candidate change"}</dt><dd>{improvement.final_candidate_delta === null ? (zh ? "未知" : "Unknown") : `${format(improvement.final_candidate_delta)} ${outcomes.metric.unit}`} · {zh ? "不代表已采用收益" : "not realized retention"}</dd></div>
+    </dl></details>
+  </section>;
+}
+
 function NativeResults({native, zh, onCitation}: {native: ResearchProgressPack["native_outcomes"]; zh: boolean; onCitation: (citation: AssessmentCitation, origin?: HTMLButtonElement) => void}) {
   const outcomes = native?.projection;
   if (!native || !outcomes) return <p className="research-native-empty">{zh ? "原生测量 · 未采集" : "Native measurements · not captured"}</p>;
@@ -54,7 +76,7 @@ function NativeResults({native, zh, onCitation}: {native: ResearchProgressPack["
       <p>{native.artifact}<br/><code>{native.artifact_sha256}</code></p>
       <dl>{points.map(([key, cn, en]) => {
         const point = outcomes[key];
-        return point ? <div key={key}><dt>{zh ? cn : en}</dt><dd><code>{point.candidate_ref}</code><br/>{zh ? "提交" : "Submitted"}: {point.submitted_at}<br/>{zh ? "收到结果" : "Result received"}: {point.observed_at}</dd></div> : null;
+        return point ? <div key={key}><dt>{zh ? cn : en}</dt><dd><code>{point.candidate_ref}</code><br/>{zh ? "提交" : "Submitted"}: {point.submitted_at ?? (zh ? "未知" : "Unknown")}<br/>{zh ? "收到结果" : "Result received"}: {point.observed_at}</dd></div> : null;
       })}</dl>
       <ul>{outcomes.limitations.map((line, i) => <li key={i}>{line}</li>)}</ul>
     </details>
@@ -81,6 +103,7 @@ export function ResearchProgress({pack: p, packs, selectedId, error, zh, onSelec
       : (zh ? "所选来源暂不可读，未显示其他范围的结果。" : "The selected source could not be read. No result from another scope is shown.")} {error}</p> : null}
     {p ? <>
       {p.applicability === "historical" ? <p className="research-progress-warning" role="status">{zh ? "历史分析：来源或方法已变化，不能当作当前判断。" : "Historical analysis: the source or method changed; these are not current judgments."}</p> : null}
+      {p.native_outcomes ? <RetainedImprovement native={p.native_outcomes} zh={zh} onCitation={onCitation}/> : null}
       {result ? <div className="research-process-counts">
         <div className="research-effect-metrics" key={p.pack_id}>
           {TRAJECTORY_METRICS.map(metric => {
@@ -114,13 +137,13 @@ export function ResearchProgress({pack: p, packs, selectedId, error, zh, onSelec
           })}
         </div>
       </div> : <p className="research-effect-empty" role="status">{zh ? "所选范围尚未分析" : "No analysis for the selected scope"}</p>}
-      {p.native_outcomes ? <NativeResults native={p.native_outcomes} zh={zh} onCitation={onCitation}/> : null}
       <details className="research-progress-basis"><summary>{zh ? "来源与用量" : "Sources and usage"}<span>{p.native_outcomes ? "" : (zh ? " · 原生测量未采集" : " · Native measurements uncaptured")}{zh ? " · 完整成本未知" : " · Total cost unknown"}</span></summary><dl>
         <dt>{zh ? "题目 / 运行" : "Case / run"}</dt><dd>{p.scope.case_id}<br/>{p.scope.run_id}</dd><dt>{zh ? "证据窗口" : "Evidence window"}</dt><dd>{p.scope.window}</dd><dt>{zh ? "取样规则" : "Selection rule"}</dt><dd>{p.selection.rule}</dd>
         <dt>{zh ? "总体研究片段数" : "Population of episodes"}</dt><dd>{p.selection.population_episodes == null ? (zh ? "未知" : "Unknown") : number(p.selection.population_episodes)}</dd>
         {result ? <><dt>{zh ? "分析方法" : "Analysis method"}</dt><dd>{result.method_version} · {zh ? "Agent 判读未校准" : "Uncalibrated Agent judgment"}</dd><dt>{zh ? "结果标识" : "Result key"}</dt><dd><code>{result.analysis_key}</code></dd><dt>{zh ? "分析限制" : "Limitations"}</dt><dd>{result.limitations.map((line, i) => <p key={i}>{line}</p>)}</dd></> : null}
         <dt>{zh ? "来源版本" : "Source revision"}</dt><dd><code>{p.scope.revision}</code></dd>
       </dl>
+      {p.native_outcomes ? <NativeResults native={p.native_outcomes} zh={zh} onCitation={onCitation}/> : null}
       <section className="research-usage"><h4>{zh ? "用量与成本" : "Usage and cost"}{measured?.applicability === "historical" ? (zh ? " · 历史捕获" : " · Historical capture") : ""}</h4>
         <p>{zh ? "只读捕获，窗口可能与以上判读片段及下方 Explore 历史不同。" : "Read-only capture; its window can differ from the judgments above and Explore history below."}</p>
         <dl className="research-usage-basis">

@@ -233,6 +233,10 @@ contiguous submission indices starting at one. An invalid record uses
 `valid:false,value:null`; zero is a valid measured value, never a missing value.
 `retained_result_id:null` means retention is unknown, not that the mathematically
 best value was adopted.
+An archived measurement may explicitly use `submitted_at:null` when only its
+readback time is known; `observed_at` must still be inside the capture. This does
+not invent an original submission time, a full evaluation duration or prefix
+coverage. Known submission times retain all existing calendar/window checks.
 
 The existing typed benchmark-toolkit prepare owner returns a separate bounded
 `native_outcomes` projection: first valid submission-one value when the prefix
@@ -258,8 +262,53 @@ Synthetic packaged native readback (not benchmark findings):
 ![Native outcome source disclosure](../development/assets/research-analysis/native-outcomes-desktop.png)
 
 [Narrow-screen source disclosure](../development/assets/research-analysis/native-outcomes-mobile.png).
-The existing six-judgment first viewport is unchanged; these facts live in its
-existing source disclosure, with costs in their separate disclosure.
+Detailed native facts live in the existing source disclosure, with costs in
+their separate disclosure.
+
+#### Quantified retained improvement
+
+To expose observed outcome value alongside the six process judgments, the same
+provider may optionally add `comparison` to its native capture:
+
+```json
+"comparison": {
+  "baseline_result_id": "result-1",
+  "retention_citation": {"source_id":"adoption-source", "first_line":4, "last_line":8}
+}
+```
+
+The provider fixes a comparable baseline under the same native objective,
+metric, unit and evaluation conditions; it must not select a baseline after
+seeing the gain. Its result must be valid and present in this same scoped
+capture, no later in submission order than the retained result. An incomplete
+prefix may still contain an explicitly known baseline; that does not establish
+the study's initial value. The retention citation must bind original evidence
+of actual retained source/result identity, not a proposed decision or the
+mathematically best score. Source binding checks provenance, not the truth or
+comparability of provider claims.
+
+The TypeScript owner adds `retained_improvement`: baseline, retention citation,
+signed `value` and separate `final_candidate_delta`. Higher-is-better uses
+retained minus baseline; lower-is-better reverses the subtraction. For example,
+native error 7 → retained 3 yields **+4 points**. Retaining 7 after rejecting a
+candidate with error 9 yields **0**, with candidate delta **−2**; avoiding that
+regression is not a new realized benefit. An actually retained regression is
+negative. Baseline or retention may explicitly be `null`; missing baseline,
+retained result or retention evidence leaves realized value `null`, never zero.
+Invalid references, unbound citations and non-finite differences are rejected.
+Omitting `comparison` preserves the previous projection shape and does not
+invent a baseline from the first/best result. Native-only registration still
+does not execute an analysis or alter its key.
+
+The Goal shows one compact **Retained improvement / 已采用成果增量** value before
+the process judgments. Hover, keyboard focus or tap explains the metric;
+expansion exposes both measured endpoints, retention evidence and candidate
+change. Historical/unreadable provenance remains visible. The value describes
+the declared native comparison, not objective completion, a general utility
+score, matched LoopX uplift or causal attribution. Do not add across tasks,
+units, overlapping windows or inherited results. Displayed precision is rounded;
+original endpoint values remain readable in their citations. Pack selection
+creates browser history so Back/Forward restores its scope and evidence.
 
 ### Solver capture registration