eduevidence 6.2.0 → 6.3.0

package/examples/workplace-ai-assistant/reports-5themes/EduEvidence_Report_academic.html

@@ -1911,7 +1911,7 @@ select:focus-visible,
 <button type="button" class="report-view-btn" data-report-view="full" aria-pressed="false">完整报告</button>
 </nav>
 </header>
-<div class="report-page report-page-brief" data-report-page="brief"><nav class="brief-navigation" aria-label="摘要导航"><a href="#brief-decision-zh"><span>01</span>先看结论</a><a href="#brief-outcomes-zh"><span>02</span>任务表现 ≠ 学习效果</a><a href="#brief-tribunal-zh"><span>03</span>证据裁决</a><a href="#brief-action-zh"><span>04</span>从证据到行动</a><a href="#brief-sources-zh"><span>05</span>关键来源</a></nav><div class="brief-reading-content"><section class="brief-block brief-decision" id="brief-decision-zh"><header class="brief-block-header"><h2>先看结论</h2><p>该不该做、置信度多高、最关键的证据边界在哪。</p></header><div class="brief-block-body">
+<div class="report-page report-page-brief" data-report-page="brief"><nav class="brief-navigation" aria-label="摘要导航"><a href="#brief-decision-zh"><span>01</span>先看结论</a><a href="#brief-outcomes-zh"><span>02</span>过程产出 ≠ 政策效果</a><a href="#brief-tribunal-zh"><span>03</span>证据裁决</a><a href="#brief-action-zh"><span>04</span>从证据到行动</a><a href="#brief-sources-zh"><span>05</span>关键来源</a></nav><div class="brief-reading-content"><section class="brief-block brief-decision" id="brief-decision-zh"><header class="brief-block-header"><h2>先看结论</h2><p>该不该做、置信度多高、最关键的证据边界在哪。</p></header><div class="brief-block-body">
 <div class="decision-hero pilot" data-visual="decision-hero">
   <div class="hero-decision">
     <span class="eyebrow">建议决策</span>
@@ -1926,7 +1926,7 @@ select:focus-visible,
     <article class="hero-insight next"><span>下一步</span><p class="hero-insight-text">在已批准知识库上开展人工监督试点，每条回复均经坐席复核，并跟踪升级率与纠错率。</p></article>
   </div>
   <p class="hero-provenance"><span>证据 / 来源</span> · 4 / 3</p>
-</div></div></section><section class="brief-block brief-outcomes" id="brief-outcomes-zh"><header class="brief-block-header"><h2>任务表现 ≠ 学习效果</h2><p>只展示真正有解释力的结果分离；正向、负向与零效应按 effect_direction 编码。</p></header><div class="brief-block-body"><div class="outcome-separation" data-visual="outcome-separation"><div class="visual-heading"><h3>结果分离 · 任务表现 ≠ 学习效果</h3><p>将不同结果类型分开裁决，避免把训练时更快、更高分直接等同于真正学会。</p><p class="semantic-note">效应方向来自 evidence.effect_direction；“支持某个主张”不等于“结果是正向”。</p></div><div class="outcome-groups"><article class="outcome-group outcome-other"><h3>其他结果</h3><ul><li><strong>Policy Effectiveness</strong><span class="outcome-states"><span class="dir pos">正向效应 2</span></span></li><li><strong>Implementation Risk</strong><span class="outcome-states"><span class="dir neg">负向效应 2</span></span></li></ul></article></div></div></div></section><section class="brief-block brief-tribunal" id="brief-tribunal-zh"><header class="brief-block-header"><h2>证据裁决</h2><p>支持、不确定、被反驳与缺失证据分开放置，不把长段落平铺在同一层。</p></header><div class="brief-block-body"><div class="evidence-tribunal" data-visual="evidence-tribunal-grid"><p class="tribunal-summary"><strong>决策：</strong>试点验证 · <strong>置信度：</strong>中</p><div class="tribunal-grid"><article class="tribunal-card supported"><header><span aria-hidden="true">✓</span><h3>可以主张</h3><span class="tribunal-count">4</span></header><ul><li><p>在有边界的客服场景中，AI 辅助可缩短会话处理时间；吞吐量是另一项指标。 — E-001</p><div class="tribunal-evidence-refs"><code>E-001</code></div></li><li><p>ChatGPT 缩短了短篇职业写作任务用时；对客服而言属于间接证据。 — E-002</p><div class="tribunal-evidence-refs"><code>E-002</code></div></li><li><p>在超出能力边界的任务中 AI 可能降低正确率；咨询实验对客服属于间接证据。 — E-003</p><div class="tribunal-evidence-refs"><code>E-003</code></div></li></ul><details class="tribunal-more"><summary>查看其余 1 条</summary><ul><li><p>须单独监测资深高技能客服的质量，不能假定所有岗位群体均受益。 — E-004</p><div class="tribunal-evidence-refs"><code>E-004</code></div></li></ul></details></article><article class="tribunal-card uncertain"><header><span aria-hidden="true">?</span><h3>尚不能主张</h3><span class="tribunal-count">1</span></header><ul><li><p>这组选编证据尚未确立本企业净成本、隐私事件率、亚组服务质量及持续表现 [无直接证据]。</p></li></ul></article><article class="tribunal-card contradicted"><header><span aria-hidden="true">×</span><h3>被反驳的主张</h3><span class="tribunal-count">0</span></header><ul><li>无数据。</li></ul></article><article class="tribunal-card missing"><header><span aria-hidden="true">…</span><h3>缺失证据</h3><span class="tribunal-count">1</span></header><ul><li><p>这组选编证据尚未确立本企业净成本、隐私事件率、亚组服务质量及持续表现 [无直接证据]。</p></li></ul></article></div></div></div></section><section class="brief-block brief-action" id="brief-action-zh"><header class="brief-block-header"><h2>从证据到行动</h2><p>适用性、护栏、停止条件与评价连成一条可执行路径。</p></header><div class="brief-block-body"><div class="evidence-to-action" data-visual="evidence-to-action"><article class="action-node action-evidence"><span>证据</span><p class="action-node-text">在有边界的客服场景中，AI 辅助可缩短会话处理时间；吞吐量是另一项指标。 — E-001</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-applicability"><span>适用性</span><p class="action-node-text">从低风险且知识库覆盖的队列开始；客服审查并编辑每条建议回复。; 按资历和基线技能分层；保留专家判断并记录复核工作量。; 试点前落实客户数据最小化与脱敏、访问和留存限制，并审查供应商数据使用条款。; 陌生、模糊、敏感或高风险请求交由合格人员处理；禁止自动作出承诺。</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-decision"><span>决策</span><p class="action-node-text">试点验证</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-guardrails"><span>护栏</span><p class="action-node-text">从低风险且知识库覆盖的队列开始；客服审查并编辑每条建议回复。</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-stop"><span>停止条件</span><p class="action-node-text">一旦核实隐私泄漏、严重不安全建议或越权客户承诺，立即暂停，调查后再决定是否恢复。; 若盲评质量越过预先约定的非劣界值（包括各资历分层），暂停扩展。</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-evaluation"><span>评价</span><p class="action-node-text">仅在质量非劣、每付薪工时有效解决量提高、净成本可接受且无未解决严重安全事件时考虑扩展；阈值须在本地预先约定。</p></article></div></div></section><section class="brief-block brief-lieflat" id="brief-lieflat-zh"><header class="brief-block-header"><h2>Lieflat 实证手作画廊</h2><p>AI 按数据形状从 Lieflat 目录选型编排；每张图的数字都可溯源到 result.json。</p></header><div class="brief-block-body"><div class="lieflat-gallery-container"><figure class="lieflat-card" data-lieflat data-visual="lieflat-bubble_almanac" data-chart-id="lieflat-bubble-almanac.svg"><h3 class="lieflat-title">发表年份 × 结果维度文献年历</h3><p class="lieflat-sub">气泡面积 ∝ 该格研究数 · 实心圆 = 有显著结果</p><div class="lieflat-figure"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 540 300" width="100%" height="100%" role="img" aria-label="发表年份 × 结果维度文献年历" style="background:#FAFAFA;">
+</div></div></section><section class="brief-block brief-outcomes" id="brief-outcomes-zh"><header class="brief-block-header"><h2>过程产出 ≠ 政策效果</h2><p>只展示真正有解释力的结果分离；正向、负向与零效应按 effect_direction 编码。</p></header><div class="brief-block-body"><div class="outcome-separation" data-visual="outcome-separation"><div class="visual-heading"><h3>结果分离 · 过程产出 ≠ 政策效果</h3><p>将不同结果类型分开裁决，避免把试点期产出或效率提升直接等同于真实政策效果。</p><p class="semantic-note">效应方向来自 evidence.effect_direction；“支持某个主张”不等于“结果是正向”。</p></div><div class="outcome-groups"><article class="outcome-group outcome-learning"><h3>效果 / 目标结果</h3><ul><li><strong>政策有效性</strong><span class="outcome-states"><span class="dir pos">正向效应 2</span></span></li></ul></article><article class="outcome-group outcome-risk"><h3>风险 / 实施风险</h3><ul><li><strong>实施风险</strong><span class="outcome-states"><span class="dir neg">负向效应 2</span></span></li></ul></article></div></div></div></section><section class="brief-block brief-tribunal" id="brief-tribunal-zh"><header class="brief-block-header"><h2>证据裁决</h2><p>支持、不确定、被反驳与缺失证据分开放置，不把长段落平铺在同一层。</p></header><div class="brief-block-body"><div class="evidence-tribunal" data-visual="evidence-tribunal-grid"><p class="tribunal-summary"><strong>决策：</strong>试点验证 · <strong>置信度：</strong>中</p><div class="tribunal-grid"><article class="tribunal-card supported"><header><span aria-hidden="true">✓</span><h3>可以主张</h3><span class="tribunal-count">4</span></header><ul><li><p>在有边界的客服场景中，AI 辅助可缩短会话处理时间；吞吐量是另一项指标。 — E-001</p><div class="tribunal-evidence-refs"><code>E-001</code></div></li><li><p>ChatGPT 缩短了短篇职业写作任务用时；对客服而言属于间接证据。 — E-002</p><div class="tribunal-evidence-refs"><code>E-002</code></div></li><li><p>在超出能力边界的任务中 AI 可能降低正确率；咨询实验对客服属于间接证据。 — E-003</p><div class="tribunal-evidence-refs"><code>E-003</code></div></li></ul><details class="tribunal-more"><summary>查看其余 1 条</summary><ul><li><p>须单独监测资深高技能客服的质量，不能假定所有岗位群体均受益。 — E-004</p><div class="tribunal-evidence-refs"><code>E-004</code></div></li></ul></details></article><article class="tribunal-card uncertain"><header><span aria-hidden="true">?</span><h3>尚不能主张</h3><span class="tribunal-count">1</span></header><ul><li><p>这组选编证据尚未确立本企业净成本、隐私事件率、亚组服务质量及持续表现 [无直接证据]。</p></li></ul></article><article class="tribunal-card contradicted"><header><span aria-hidden="true">×</span><h3>被反驳的主张</h3><span class="tribunal-count">0</span></header><ul><li>无数据。</li></ul></article><article class="tribunal-card missing"><header><span aria-hidden="true">…</span><h3>缺失证据</h3><span class="tribunal-count">1</span></header><ul><li><p>这组选编证据尚未确立本企业净成本、隐私事件率、亚组服务质量及持续表现 [无直接证据]。</p></li></ul></article></div></div></div></section><section class="brief-block brief-action" id="brief-action-zh"><header class="brief-block-header"><h2>从证据到行动</h2><p>适用性、护栏、停止条件与评价连成一条可执行路径。</p></header><div class="brief-block-body"><div class="evidence-to-action" data-visual="evidence-to-action"><article class="action-node action-evidence"><span>证据</span><p class="action-node-text">在有边界的客服场景中，AI 辅助可缩短会话处理时间；吞吐量是另一项指标。 — E-001</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-applicability"><span>适用性</span><p class="action-node-text">从低风险且知识库覆盖的队列开始；客服审查并编辑每条建议回复。; 按资历和基线技能分层；保留专家判断并记录复核工作量。; 试点前落实客户数据最小化与脱敏、访问和留存限制，并审查供应商数据使用条款。; 陌生、模糊、敏感或高风险请求交由合格人员处理；禁止自动作出承诺。</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-decision"><span>决策</span><p class="action-node-text">试点验证</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-guardrails"><span>护栏</span><p class="action-node-text">从低风险且知识库覆盖的队列开始；客服审查并编辑每条建议回复。</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-stop"><span>停止条件</span><p class="action-node-text">一旦核实隐私泄漏、严重不安全建议或越权客户承诺，立即暂停，调查后再决定是否恢复。; 若盲评质量越过预先约定的非劣界值（包括各资历分层），暂停扩展。</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-evaluation"><span>评价</span><p class="action-node-text">仅在质量非劣、每付薪工时有效解决量提高、净成本可接受且无未解决严重安全事件时考虑扩展；阈值须在本地预先约定。</p></article></div></div></section><section class="brief-block brief-lieflat" id="brief-lieflat-zh"><header class="brief-block-header"><h2>Lieflat 实证手作画廊</h2><p>AI 按数据形状从 Lieflat 目录选型编排；每张图的数字都可溯源到 result.json。</p></header><div class="brief-block-body"><div class="lieflat-gallery-container"><figure class="lieflat-card" data-lieflat data-visual="lieflat-bubble_almanac" data-chart-id="lieflat-bubble-almanac.svg"><h3 class="lieflat-title">发表年份 × 结果维度文献年历</h3><p class="lieflat-sub">气泡面积 ∝ 该格研究数 · 实心圆 = 有显著结果</p><div class="lieflat-figure"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 540 300" width="100%" height="100%" role="img" aria-label="发表年份 × 结果维度文献年历" style="background:#FAFAFA;">
 <line x1="44" y1="70.0" x2="520" y2="70.0" stroke="#E2E8F0" stroke-width="0.5" class="lf-fade" style="--motion-delay:0ms"/>
 <line x1="44" y1="77.0" x2="520" y2="77.0" stroke="#E2E8F0" stroke-width="0.5" class="lf-fade" style="--motion-delay:14ms"/>
 <line x1="44" y1="84.0" x2="520" y2="84.0" stroke="#E2E8F0" stroke-width="0.5" class="lf-fade" style="--motion-delay:28ms"/>
@@ -1955,27 +1955,27 @@ select:focus-visible,
 <line x1="44" y1="245.0" x2="520" y2="245.0" stroke="#E2E8F0" stroke-width="0.5" class="lf-fade" style="--motion-delay:350ms"/>
 <line x1="44" y1="252.0" x2="520" y2="252.0" stroke="#E2E8F0" stroke-width="0.5" class="lf-fade" style="--motion-delay:364ms"/>
 <line x1="44" y1="259.0" x2="520" y2="259.0" stroke="#E2E8F0" stroke-width="0.5" class="lf-fade" style="--motion-delay:378ms"/>
-<text x="150" y="76" fill="#0F172A" font-size="9" font-weight="600" text-anchor="middle" class="lf-fade" style="--motion-delay:40ms">policy_effecti</text>
-<text x="484.8" y="76" fill="#0F172A" font-size="9" font-weight="600" text-anchor="middle" class="lf-fade" style="--motion-delay:140ms">implementation</text>
+<text x="150" y="76" fill="#0F172A" font-size="9" font-weight="600" text-anchor="middle" class="lf-fade" style="--motion-delay:40ms">政策有效性</text>
+<text x="514.0" y="76" fill="#0F172A" font-size="9" font-weight="600" text-anchor="middle" class="lf-fade" style="--motion-delay:140ms">实施风险</text>
 <text x="96" y="96.0" fill="#475569" font-size="9" font-weight="800" text-anchor="end" class="lf-fade" style="--motion-delay:80ms">2023</text>
 <text x="96" y="140.0" fill="#475569" font-size="9" font-weight="800" text-anchor="end" class="lf-fade" style="--motion-delay:180ms">2025</text>
 <text x="96" y="184.0" fill="#475569" font-size="9" font-weight="800" text-anchor="end" class="lf-fade" style="--motion-delay:280ms">2026</text>
-<circle cx="150.0" cy="92.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:100ms"><title>policy_effectiveness (2023) — N = 1 篇研究, 显著 = 0</title></circle>
-<circle cx="150.0" cy="136.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:112ms"><title>policy_effectiveness (2025) — N = 1 篇研究, 显著 = 0</title></circle>
-<circle cx="520.0" cy="136.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:124ms"><title>implementation_risk (2025) — N = 1 篇研究, 显著 = 0</title></circle>
-<circle cx="520.0" cy="180.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:136ms"><title>implementation_risk (2026) — N = 1 篇研究, 显著 = 0</title></circle>
+<circle cx="150.0" cy="92.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:100ms"><title>政策有效性 (2023) — N = 1 篇研究, 显著 = 0</title></circle>
+<circle cx="150.0" cy="136.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:112ms"><title>政策有效性 (2025) — N = 1 篇研究, 显著 = 0</title></circle>
+<circle cx="520.0" cy="136.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:124ms"><title>实施风险 (2025) — N = 1 篇研究, 显著 = 0</title></circle>
+<circle cx="520.0" cy="180.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:136ms"><title>实施风险 (2026) — N = 1 篇研究, 显著 = 0</title></circle>
 </svg></div><figcaption class="lieflat-caption">仅当证据集携带发表年份与结果维度时绘制。</figcaption><p class="lieflat-src">L9 Bubble Almanac · Evidence.year X Dimension</p></figure><figure class="lieflat-card" data-lieflat data-visual="lieflat-matrix_heat" data-chart-id="lieflat-matrix-heat.svg"><h3 class="lieflat-title">年份 × 结果维度证据密度</h3><p class="lieflat-sub">每格数字 = 该年份该结果维度的证据条数</p><div class="lieflat-figure"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 540 200" width="100%" height="100%" role="img" aria-label="年份 × 结果维度证据密度" style="background:#FAFAFA;">
 <text x="211.7" y="66" text-anchor="middle" font-size="9" font-weight="800" fill="#475569" class="lf-fade" style="--motion-delay:40ms">2023</text>
 <text x="335.0" y="66" text-anchor="middle" font-size="9" font-weight="800" fill="#475569" class="lf-fade" style="--motion-delay:140ms">2025</text>
 <text x="458.3" y="66" text-anchor="middle" font-size="9" font-weight="800" fill="#475569" class="lf-fade" style="--motion-delay:240ms">2026</text>
-<text x="138" y="97.0" text-anchor="end" font-size="8" font-weight="600" fill="#475569" class="lf-fade" style="--motion-delay:60ms">policy effecti</text>
+<text x="138" y="97.0" text-anchor="end" font-size="8" font-weight="600" fill="#475569" class="lf-fade" style="--motion-delay:60ms">政策有效性</text>
 <rect x="152.0" y="80.0" width="119.3" height="28" rx="4" fill="#0F172A" fill-opacity="0.90" class="lf-pop" style="--motion-delay:0ms"/>
 <text x="211.7" y="97.0" text-anchor="middle" font-size="8.0" font-weight="800" fill="#0F172A" class="lf-pop" style="--motion-delay:60ms">1</text>
 <rect x="275.3" y="80.0" width="119.3" height="28" rx="4" fill="#0F172A" fill-opacity="0.90" class="lf-pop" style="--motion-delay:12ms"/>
 <text x="335.0" y="97.0" text-anchor="middle" font-size="8.0" font-weight="800" fill="#0F172A" class="lf-pop" style="--motion-delay:72ms">1</text>
 <rect x="398.7" y="80.0" width="119.3" height="28" rx="4" fill="#E2E8F0" fill-opacity="0.5" class="lf-fade" style="--motion-delay:24ms"/>
 <text x="458.3" y="97.0" text-anchor="middle" font-size="8.0" font-weight="800" fill="#0F172A" class="lf-pop" style="--motion-delay:84ms">0</text>
-<text x="138" y="129.0" text-anchor="end" font-size="8" font-weight="600" fill="#475569" class="lf-fade" style="--motion-delay:160ms">implementation</text>
+<text x="138" y="129.0" text-anchor="end" font-size="8" font-weight="600" fill="#475569" class="lf-fade" style="--motion-delay:160ms">实施风险</text>
 <rect x="152.0" y="112.0" width="119.3" height="28" rx="4" fill="#E2E8F0" fill-opacity="0.5" class="lf-fade" style="--motion-delay:12ms"/>
 <text x="211.7" y="129.0" text-anchor="middle" font-size="8.0" font-weight="800" fill="#0F172A" class="lf-pop" style="--motion-delay:72ms">0</text>
 <rect x="275.3" y="112.0" width="119.3" height="28" rx="4" fill="#0F172A" fill-opacity="0.90" class="lf-pop" style="--motion-delay:24ms"/>
@@ -2132,7 +2132,7 @@ select:focus-visible,
 </svg></div><figcaption class="lieflat-caption">让读者一眼看出方法学短板集中在哪些检查项。</figcaption><p class="lieflat-src">L15 Ballot Tally · Methodology.flag Rates</p></figure></div></div></section><section class="brief-block brief-sources" id="brief-sources-zh"><header class="brief-block-header"><h2>关键来源</h2><p>摘要页只列最关键的来源；完整溯源在完整报告中展开。</p></header><div class="brief-block-body"><div class="brief-source-grid"><article class="brief-source"><code>S-001</code><h3><a href="https://academic.oup.com/qje/article/140/2/889/7990658">Generative AI at Work</a></h3><p>T1 DOI 可验证论文 · 2025</p></article><article class="brief-source"><code>S-002</code><h3><a href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf">Experimental evidence on the productivity effects of generative artificial intelligence</a></h3><p>T1 DOI 可验证论文 · 2023</p></article><article class="brief-source"><code>S-003</code><h3><a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838">Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality</a></h3><p>T1 DOI 可验证论文 · 2026</p></article></div></div></section></div></div>
 <div class="report-page report-page-full" data-report-page="full" hidden>
 <div class="full-report-intro"><h2>完整报告</h2><p>结论前置：全部可追溯证据与方法学细节都在这里，关键论证位置穿插有意义的可视化，每个数字都能回查到 result.json。</p></div>
-<div class="full-report-layout"><aside class="full-report-toc" aria-label="目录"><div class="toc-head"><strong>目录</strong><button type="button" class="toc-collapse" aria-expanded="true" data-label-collapse="收起目录" data-label-expand="展开目录">收起目录</button></div><nav><a href="#full-01-decision" data-toc-target="full-01-decision" data-chapter-key="decision">01 结论、裁决与研究边界</a><a href="#full-02-evidence" data-toc-target="full-02-evidence" data-chapter-key="evidence">02 关键证据与结果分离</a><a href="#full-03-quality" data-toc-target="full-03-quality" data-chapter-key="quality">03 证据可信度、反证与方法审计</a><a href="#full-04-action" data-toc-target="full-04-action" data-chapter-key="action">04 适用范围与教学行动</a><a href="#full-05-evaluation" data-toc-target="full-05-evaluation" data-chapter-key="evaluation">05 试点设计、评估与停止条件</a><a href="#full-06-sources" data-toc-target="full-06-sources" data-chapter-key="sources">06 来源、溯源与附录</a></nav></aside><main class="full-report-content"><section id="full-01-decision" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>01 结论、裁决与研究边界</h2><p class="full-chapter-lead">先明确最终裁决与研究边界，再解释为什么。</p></header><div class="full-chapter-body">
+<div class="full-report-layout"><aside class="full-report-toc" aria-label="目录"><div class="toc-head"><strong>目录</strong><button type="button" class="toc-collapse" aria-expanded="true" data-label-collapse="收起目录" data-label-expand="展开目录">收起目录</button></div><nav><a href="#full-01-decision" data-toc-target="full-01-decision" data-chapter-key="decision">01 结论、裁决与研究边界</a><a href="#full-02-evidence" data-toc-target="full-02-evidence" data-chapter-key="evidence">02 关键证据与结果分离</a><a href="#full-03-quality" data-toc-target="full-03-quality" data-chapter-key="quality">03 证据可信度、反证与方法审计</a><a href="#full-04-action" data-toc-target="full-04-action" data-chapter-key="action">04 适用范围与政策行动</a><a href="#full-05-evaluation" data-toc-target="full-05-evaluation" data-chapter-key="evaluation">05 试点设计、评估与停止条件</a><a href="#full-06-sources" data-toc-target="full-06-sources" data-chapter-key="sources">06 来源、溯源与附录</a></nav></aside><main class="full-report-content"><section id="full-01-decision" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>01 结论、裁决与研究边界</h2><p class="full-chapter-lead">先明确最终裁决与研究边界，再解释为什么。</p></header><div class="full-chapter-body">
 <div class="decision-hero pilot" data-visual="decision-hero">
   <div class="hero-decision">
     <span class="eyebrow">建议决策</span>
@@ -2147,19 +2147,19 @@ select:focus-visible,
     <article class="hero-insight next"><span>下一步</span><p class="hero-insight-text">在已批准知识库上开展人工监督试点，每条回复均经坐席复核，并跟踪升级率与纠错率。</p></article>
   </div>
   <p class="hero-provenance"><span>证据 / 来源</span> · 4 / 3</p>
-</div><div class="scope-grid"><article class="scope-card"><h3>研究问题</h3><p class="scope-text">企业客服团队是否应引入生成式 AI 助手？</p></article><article class="scope-card"><h3>AI 干预</h3><p class="scope-text">政策名称：人工监督的客服助手；政策类型：机构改革；作用机制：已批准知识库 + 坐席复核</p></article><article class="scope-card"><h3>比较条件</h3><p class="scope-text">不提供生成式 AI 建议的现有客服流程。</p></article><article class="scope-card"><h3>结果构念</h3><p class="scope-text">主要结果：Policy Effectiveness、Implementation Risk；次要结果：Cost Effectiveness、equity、feasibility</p></article><article class="scope-card"><h3>研究范围</h3><p class="scope-text">时间范围：2023–2026；Evidence Types：准实验、随机对照试验</p></article><article class="scope-card"><h3>决策成功条件</h3><p class="scope-text">提高每付薪工时的有效解决量，同时维持服务质量、隐私和员工自主判断。</p></article></div></div></section><section id="full-02-evidence" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>02 关键证据与结果分离</h2><p class="full-chapter-lead">把任务表现、真实学习、保持与风险放在同一证据地图中，但不混为一谈。</p></header><div class="full-chapter-body"><div class="retrieval-grid"><article><h3>纳入标准</h3><ul></ul></article><article><h3>排除标准</h3><ul></ul></article></div><div class="retrieval-coverage"><h3>证据来源覆盖</h3><p><code>S-001</code> <code>S-002</code> <code>S-003</code></p><p class="retrieval-note">当前报告只展示 result 中真实存在的检索与来源信息；没有流程计数时不伪造 PRISMA / funnel 数字。</p></div><div class="outcome-separation" data-visual="outcome-separation"><div class="visual-heading"><h3>结果分离 · 任务表现 ≠ 学习效果</h3><p>将不同结果类型分开裁决，避免把训练时更快、更高分直接等同于真正学会。</p><p class="semantic-note">效应方向来自 evidence.effect_direction；“支持某个主张”不等于“结果是正向”。</p></div><div class="outcome-groups"><article class="outcome-group outcome-other"><h3>其他结果</h3><ul><li><strong>Policy Effectiveness</strong><span class="outcome-states"><span class="dir pos">正向效应 2</span></span></li><li><strong>Implementation Risk</strong><span class="outcome-states"><span class="dir neg">负向效应 2</span></span></li></ul></article></div></div><div class='table-wrap outcome-table'><table class='data-table'><thead><tr><th>结果类型</th><th>正向效应</th><th>负向效应</th><th>零效应</th><th>证据</th></tr></thead><tbody><tr><td><strong>Policy Effectiveness</strong><span class='raw-tag' title='原始标识'>policy_effectiveness</span></td><td class='num'>2</td><td class='num'>0</td><td class='num'>0</td><td><code>E-001</code> <code>E-002</code> </td></tr><tr><td><strong>Implementation Risk</strong><span class='raw-tag' title='原始标识'>implementation_risk</span></td><td class='num'>0</td><td class='num'>2</td><td class='num'>0</td><td><code>E-003</code> <code>E-004</code> </td></tr></tbody></table></div><div class='matrix-controls'><div class='matrix-tools'><input id='matrix-search-full-zh' type='search' placeholder='搜索证据…' aria-label='筛选 / 搜索证据'><select id='matrix-direction-full-zh' aria-label='按效应方向筛选'><option value=''>全部效应</option><option value='positive'>正向效应</option><option value='negative'>负向效应</option><option value='null'>零效应</option></select><select id='matrix-outcome-full-zh' aria-label='按结果类型筛选'><option value=''>全部结果</option><option value='implementation_risk'>Implementation Risk</option><option value='policy_effectiveness'>Policy Effectiveness</option></select></div></div><div class='table-wrap matrix-wrap'><table id='evidence-matrix-full-zh' class='data-table evidence-matrix'><thead><tr><th>ID</th><th>结果</th><th>效应</th><th>质量</th><th>主张</th><th>来源</th></tr></thead><tbody><tr data-effect="positive" data-direction="positive" data-outcome="policy_effectiveness" data-search="e-001 st-001 sample-support-all generative ai at work 在有边界的客服场景中，ai 辅助可缩短会话处理时间；吞吐量是另一项指标。 客服人员   positive s-001"><td><code>E-001</code></td><td><strong>Policy Effectiveness</strong></td><td><span class="dir pos">正向效应</span><span class="relation-note">与主张关系: 支持</span></td><td><div class="quality-cell"><span class="num">9</span><span class="quality-meter" aria-hidden="true"><i style="width:90%"></i></span></div></td><td class="claim-cell"><p>在有边界的客服场景中，AI 辅助可缩短会话处理时间；吞吐量是另一项指标。</p><details class="matrix-row-detail"><summary>查看完整证据</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>研究 ID</dt><dd>ST-001</dd></div><div class="evidence-detail-row"><dt>样本 ID</dt><dd>SAMPLE-support-all</dd></div><div class="evidence-detail-row"><dt>研究标题</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>来源标题</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>年份</dt><dd>2025</dd></div><div class="evidence-detail-row"><dt>研究设计</dt><dd>准实验</dd></div><div class="evidence-detail-row"><dt>研究人群</dt><dd>客服人员</dd></div><div class="evidence-detail-row"><dt>样本量</dt><dd>5172</dd></div><div class="evidence-detail-row"><dt>结果测量</dt><dd>会话处理时间；另列吞吐量摘要：每小时解决问题数约增加 15%。</dd></div><div class="evidence-detail-row"><dt>效应方向</dt><dd>正向效应</dd></div><div class="evidence-detail-row"><dt>与主张关系</dt><dd>支持</dd></div><div class="evidence-detail-row"><dt>局限</dt><dd>单一企业、单一工具、非随机分批上线；因果解释依赖识别假设。</dd></div><div class="evidence-detail-row"><dt>质量维度</dt><dd>D1 研究设计=1 · D2 样本质量=2 · D3 测量效度=2 · D4 时间强度=2 · D5 直接性=2</dd></div><div class="evidence-detail-row"><dt>质量分</dt><dd>9.0</dd></div><div class="evidence-detail-row"><dt>证据状态</dt><dd>已支持</dd></div><div class="evidence-detail-row"><dt>完整主张</dt><dd>在有边界的客服场景中，AI 辅助可缩短会话处理时间；吞吐量是另一项指标。</dd></div><div class="evidence-detail-row"><dt>来源位置</dt><dd>https://academic.oup.com/qje/article/140/2/889/7990658</dd></div><div class="evidence-detail-row"><dt>可验证链接</dt><dd><a href="https://academic.oup.com/qje/article/140/2/889/7990658">https://academic.oup.com/qje/article/140/2/889/7990658</a></dd></div></dl></details></td><td><a class="source-link" href="https://academic.oup.com/qje/article/140/2/889/7990658" title="Generative AI at Work"><code>S-001</code></a></td></tr><tr data-effect="positive" data-direction="positive" data-outcome="policy_effectiveness" data-search="e-002 st-002 sample-writing experimental evidence on the productivity effects of generative artificial intelligence chatgpt 缩短了短篇职业写作任务用时；对客服而言属于间接证据。 受过大学教育的职场人士   positive s-002"><td><code>E-002</code></td><td><strong>Policy Effectiveness</strong></td><td><span class="dir pos">正向效应</span><span class="relation-note">与主张关系: 支持</span></td><td><div class="quality-cell"><span class="num">7</span><span class="quality-meter" aria-hidden="true"><i style="width:70%"></i></span></div></td><td class="claim-cell"><p>ChatGPT 缩短了短篇职业写作任务用时；对客服而言属于间接证据。</p><details class="matrix-row-detail"><summary>查看完整证据</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>研究 ID</dt><dd>ST-002</dd></div><div class="evidence-detail-row"><dt>样本 ID</dt><dd>SAMPLE-writing</dd></div><div class="evidence-detail-row"><dt>研究标题</dt><dd>Experimental evidence on the productivity effects of generative artificial intelligence</dd></div><div class="evidence-detail-row"><dt>来源标题</dt><dd>Experimental evidence on the productivity effects of generative artificial intelligence</dd></div><div class="evidence-detail-row"><dt>年份</dt><dd>2023</dd></div><div class="evidence-detail-row"><dt>研究设计</dt><dd>随机对照试验</dd></div><div class="evidence-detail-row"><dt>研究人群</dt><dd>受过大学教育的职场人士</dd></div><div class="evidence-detail-row"><dt>样本量</dt><dd>453</dd></div><div class="evidence-detail-row"><dt>结果测量</dt><dd>自报任务用时约减少 40%；评定写作质量约提高 18% 是另一项指标。</dd></div><div class="evidence-detail-row"><dt>效应方向</dt><dd>正向效应</dd></div><div class="evidence-detail-row"><dt>与主张关系</dt><dd>支持</dd></div><div class="evidence-detail-row"><dt>局限</dt><dd>间接证据，不可直接推广：短时激励写作任务不要求精确事实或客户特定情境。</dd></div><div class="evidence-detail-row"><dt>质量维度</dt><dd>D1 研究设计=2 · D2 样本质量=2 · D3 测量效度=1 · D4 时间强度=1 · D5 直接性=1</dd></div><div class="evidence-detail-row"><dt>质量分</dt><dd>7.0</dd></div><div class="evidence-detail-row"><dt>证据状态</dt><dd>已支持</dd></div><div class="evidence-detail-row"><dt>完整主张</dt><dd>ChatGPT 缩短了短篇职业写作任务用时；对客服而言属于间接证据。</dd></div><div class="evidence-detail-row"><dt>来源位置</dt><dd>https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</dd></div><div class="evidence-detail-row"><dt>可验证链接</dt><dd><a href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf">https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</a></dd></div></dl></details></td><td><a class="source-link" href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf" title="Experimental evidence on the productivity effects of generative artificial intelligence"><code>S-002</code></a></td></tr><tr data-effect="negative" data-direction="negative" data-outcome="implementation_risk" data-search="e-003 st-003 sample-consulting-outside navigating the jagged technological frontier: field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality 在超出能力边界的任务中 ai 可能降低正确率；咨询实验对客服属于间接证据。 参与超出能力边界实验的 bcg 顾问   negative s-003"><td><code>E-003</code></td><td><strong>Implementation Risk</strong></td><td><span class="dir neg">负向效应</span><span class="relation-note">与主张关系: 支持</span></td><td><div class="quality-cell"><span class="num">8</span><span class="quality-meter" aria-hidden="true"><i style="width:80%"></i></span></div></td><td class="claim-cell"><p>在超出能力边界的任务中 AI 可能降低正确率；咨询实验对客服属于间接证据。</p><details class="matrix-row-detail"><summary>查看完整证据</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>研究 ID</dt><dd>ST-003</dd></div><div class="evidence-detail-row"><dt>样本 ID</dt><dd>SAMPLE-consulting-outside</dd></div><div class="evidence-detail-row"><dt>研究标题</dt><dd>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality</dd></div><div class="evidence-detail-row"><dt>来源标题</dt><dd>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality</dd></div><div class="evidence-detail-row"><dt>年份</dt><dd>2026</dd></div><div class="evidence-detail-row"><dt>研究设计</dt><dd>随机对照试验</dd></div><div class="evidence-detail-row"><dt>研究人群</dt><dd>参与超出能力边界实验的 BCG 顾问</dd></div><div class="evidence-detail-row"><dt>样本量</dt><dd>373</dd></div><div class="evidence-detail-row"><dt>结果测量</dt><dd>商业建议正确率：合并 AI 组约低 19 个百分点；表 7 为 373 人，不是 758 人。</dd></div><div class="evidence-detail-row"><dt>效应方向</dt><dd>负向效应</dd></div><div class="evidence-detail-row"><dt>与主张关系</dt><dd>支持</dd></div><div class="evidence-detail-row"><dt>局限</dt><dd>间接证据，不可直接推广：咨询顾问处理实验商业案例，并非真实客服工单。</dd></div><div class="evidence-detail-row"><dt>质量维度</dt><dd>D1 研究设计=2 · D2 样本质量=2 · D3 测量效度=2 · D4 时间强度=1 · D5 直接性=1</dd></div><div class="evidence-detail-row"><dt>质量分</dt><dd>8.0</dd></div><div class="evidence-detail-row"><dt>证据状态</dt><dd>已支持</dd></div><div class="evidence-detail-row"><dt>完整主张</dt><dd>在超出能力边界的任务中 AI 可能降低正确率；咨询实验对客服属于间接证据。</dd></div><div class="evidence-detail-row"><dt>来源位置</dt><dd>https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</dd></div><div class="evidence-detail-row"><dt>可验证链接</dt><dd><a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838">https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</a></dd></div></dl></details></td><td><a class="source-link" href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838" title="Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality"><code>S-003</code></a></td></tr><tr data-effect="negative" data-direction="negative" data-outcome="implementation_risk" data-search="e-004 st-001 sample-support-all generative ai at work 须单独监测资深高技能客服的质量，不能假定所有岗位群体均受益。 资深高技能客服人员   negative s-001"><td><code>E-004</code></td><td><strong>Implementation Risk</strong></td><td><span class="dir neg">负向效应</span><span class="relation-note">与主张关系: 支持</span></td><td><div class="quality-cell"><span class="num">8</span><span class="quality-meter" aria-hidden="true"><i style="width:80%"></i></span></div></td><td class="claim-cell"><p>须单独监测资深高技能客服的质量，不能假定所有岗位群体均受益。</p><details class="matrix-row-detail"><summary>查看完整证据</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>研究 ID</dt><dd>ST-001</dd></div><div class="evidence-detail-row"><dt>样本 ID</dt><dd>SAMPLE-support-all</dd></div><div class="evidence-detail-row"><dt>研究标题</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>来源标题</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>年份</dt><dd>2025</dd></div><div class="evidence-detail-row"><dt>研究设计</dt><dd>准实验</dd></div><div class="evidence-detail-row"><dt>研究人群</dt><dd>资深高技能客服人员</dd></div><div class="evidence-detail-row"><dt>结果测量</dt><dd>资历最深、技能最高的客服群体出现小幅质量下降。</dd></div><div class="evidence-detail-row"><dt>效应方向</dt><dd>负向效应</dd></div><div class="evidence-detail-row"><dt>与主张关系</dt><dd>支持</dd></div><div class="evidence-detail-row"><dt>局限</dt><dd>与 E-001 属于同一研究；未提取该亚组样本量，不是独立重复验证。</dd></div><div class="evidence-detail-row"><dt>质量维度</dt><dd>D1 研究设计=1 · D2 样本质量=1 · D3 测量效度=2 · D4 时间强度=2 · D5 直接性=2</dd></div><div class="evidence-detail-row"><dt>质量分</dt><dd>8.0</dd></div><div class="evidence-detail-row"><dt>证据状态</dt><dd>已支持</dd></div><div class="evidence-detail-row"><dt>完整主张</dt><dd>须单独监测资深高技能客服的质量，不能假定所有岗位群体均受益。</dd></div><div class="evidence-detail-row"><dt>来源位置</dt><dd>https://academic.oup.com/qje/article/140/2/889/7990658</dd></div><div class="evidence-detail-row"><dt>可验证链接</dt><dd><a href="https://academic.oup.com/qje/article/140/2/889/7990658">https://academic.oup.com/qje/article/140/2/889/7990658</a></dd></div></dl></details></td><td><a class="source-link" href="https://academic.oup.com/qje/article/140/2/889/7990658" title="Generative AI at Work"><code>S-001</code></a></td></tr></tbody></table></div></div></section><section id="full-03-quality" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>03 证据可信度、反证与方法审计</h2><p class="full-chapter-lead">检查证据为什么可信、哪里冲突，以及哪些结论必须降级。</p></header><div class="full-chapter-body"><section class="method-review-group"><header class="method-review-title"><h3>审查目标：overall</h3><span class="method-verdict">关注</span></header><div class="method-audit-grid"><article class="method-audit-item method-met"><div class="method-audit-head"><strong>evidence level</strong><span class="method-status">通过</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>causal identification</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>external validity</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-missing"><div class="method-audit-head"><strong>cost evidence</strong><span class="method-status">缺失</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>stakeholder representation</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>implementation evidence</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>equity analysis</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>effect size reported</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>uncertainty quantified</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-met"><div class="method-audit-head"><strong>comparator clarity</strong><span class="method-status">通过</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>publication bias risk</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-met"><div class="method-audit-head"><strong>generalizability claims</strong><span class="method-status">通过</span></div><p></p></article></div><div class="method-guard-wrap"><strong>任务 vs 学习护栏：</strong><p class="method-guard">教学结果不适用；生产率并非学生学习效果主张。</p></div></section><p>无数据。</p><div class="evidence-tribunal" data-visual="evidence-tribunal-grid"><p class="tribunal-summary"><strong>决策：</strong>试点验证 · <strong>置信度：</strong>中</p><div class="tribunal-grid"><article class="tribunal-card supported"><header><span aria-hidden="true">✓</span><h3>可以主张</h3><span class="tribunal-count">4</span></header><ul><li><p>在有边界的客服场景中，AI 辅助可缩短会话处理时间；吞吐量是另一项指标。 — E-001</p><div class="tribunal-evidence-refs"><code>E-001</code></div></li><li><p>ChatGPT 缩短了短篇职业写作任务用时；对客服而言属于间接证据。 — E-002</p><div class="tribunal-evidence-refs"><code>E-002</code></div></li><li><p>在超出能力边界的任务中 AI 可能降低正确率；咨询实验对客服属于间接证据。 — E-003</p><div class="tribunal-evidence-refs"><code>E-003</code></div></li><li><p>须单独监测资深高技能客服的质量，不能假定所有岗位群体均受益。 — E-004</p><div class="tribunal-evidence-refs"><code>E-004</code></div></li></ul></article><article class="tribunal-card uncertain"><header><span aria-hidden="true">?</span><h3>尚不能主张</h3><span class="tribunal-count">1</span></header><ul><li><p>这组选编证据尚未确立本企业净成本、隐私事件率、亚组服务质量及持续表现 [无直接证据]。</p></li></ul></article><article class="tribunal-card contradicted"><header><span aria-hidden="true">×</span><h3>被反驳的主张</h3><span class="tribunal-count">0</span></header><ul><li>无数据。</li></ul></article><article class="tribunal-card missing"><header><span aria-hidden="true">…</span><h3>缺失证据</h3><span class="tribunal-count">1</span></header><ul><li><p>这组选编证据尚未确立本企业净成本、隐私事件率、亚组服务质量及持续表现 [无直接证据]。</p></li></ul></article></div><details class="supporting-visual"><summary>EvidenceFlow 协议</summary><svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="EvidenceFlow 协议流程"><title>EvidenceFlow 协议流程</title><desc>从问题框架、检索、抓取验证、证据抽取、反方质疑、方法审计、裁决到适用性与干预评价的完整流程。</desc><defs><marker id="arr-zh-full-workflow" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">EvidenceFlow 协议</text><rect x="24" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="56.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="56.0" y="126.0">问题框架</tspan></text><line x1="88" y1="126" x2="100" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="100" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="132.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="132.0" y="126.0">检索</tspan></text><line x1="164" y1="126" x2="176" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="176" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="208.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="208.0" y="126.0">抓取</tspan><tspan x="208.0" y="140.0">验证</tspan></text><line x1="240" y1="126" x2="252" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="252" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="284.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="284.0" y="126.0">证据抽取</tspan></text><line x1="316" y1="126" x2="328" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="328" y="100" width="64" height="52" rx="8" fill="#5E8A6A"/><text x="360.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="360.0" y="126.0">反方质疑</tspan></text><line x1="392" y1="126" x2="404" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="404" y="100" width="64" height="52" rx="8" fill="#5E8A6A"/><text x="436.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="436.0" y="126.0">方法</tspan><tspan x="436.0" y="140.0">审计</tspan></text><line x1="468" y1="126" x2="480" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="480" y="100" width="64" height="52" rx="8" fill="#C99A4A"/><text x="512.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="512.0" y="126.0">裁决</tspan></text><line x1="544" y1="126" x2="556" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="556" y="100" width="64" height="52" rx="8" fill="#C99A4A"/><text x="588.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="588.0" y="126.0">适用性</tspan></text><line x1="620" y1="126" x2="632" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="632" y="100" width="64" height="52" rx="8" fill="#C99A4A"/><text x="664.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="664.0" y="126.0">干预</tspan><tspan x="664.0" y="140.0">评价</tspan></text></svg></details><details class="supporting-visual"><summary>裁决信息图</summary><svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="证据裁决信息图"><title>证据裁决信息图</title><desc>可以主张与不可主张的证据 ID 与建议决策徽章；完整主张文本见下方裁决卡片。</desc><defs><marker id="arr-zh-full-tribunal" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">证据裁决</text><text x="24" y="180" font-size="13" font-weight="700" fill="#3A3833">冲突来源</text><text x="24" y="202" font-size="11" fill="#8A867E">详见下方裁决卡片</text><rect x="24" y="216" width="180" height="28" rx="14" fill="#C99A4A"/><text x="114" y="235" text-anchor="middle" font-size="13" font-weight="700" fill="#fff">试点验证</text><text x="250" y="80" font-size="13" font-weight="700" fill="#5E8A6A">可以主张 (4)</text><circle cx="254" cy="100" r="3" fill="#5E8A6A"/><text x="266" y="105" font-size="11" fill="#3A3833">E-001</text><circle cx="254" cy="122" r="3" fill="#5E8A6A"/><text x="266" y="127" font-size="11" fill="#3A3833">E-002</text><circle cx="254" cy="144" r="3" fill="#5E8A6A"/><text x="266" y="149" font-size="11" fill="#3A3833">E-003</text><circle cx="254" cy="166" r="3" fill="#5E8A6A"/><text x="266" y="171" font-size="11" fill="#3A3833">E-004</text><text x="490" y="80" font-size="13" font-weight="700" fill="#A85B53">不可主张 (1)</text><circle cx="494" cy="100" r="3" fill="#A85B53"/><text x="506" y="105" font-size="11" fill="#3A3833">（无）</text></svg></details></div><div class="trace-chain" data-visual="trace-chain"><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-001</strong><p class="trace-claim-text">在有边界的客服场景中，AI 辅助可缩短会话处理时间；吞吐量是另一项指标。</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-001</code><span>Policy Effectiveness</span><span class="dir pos">正向效应</span><span class="trace-relation">主张关系: 支持</span><span>Q 9.0</span><span class="trace-arrow">→</span><a href="https://academic.oup.com/qje/article/140/2/889/7990658"><code>S-001</code></a></div></div></article><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-002</strong><p class="trace-claim-text">ChatGPT 缩短了短篇职业写作任务用时；对客服而言属于间接证据。</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-002</code><span>Policy Effectiveness</span><span class="dir pos">正向效应</span><span class="trace-relation">主张关系: 支持</span><span>Q 7.0</span><span class="trace-arrow">→</span><a href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf"><code>S-002</code></a></div></div></article><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-003</strong><p class="trace-claim-text">在超出能力边界的任务中 AI 可能降低正确率；咨询实验对客服属于间接证据。</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-003</code><span>Implementation Risk</span><span class="dir neg">负向效应</span><span class="trace-relation">主张关系: 支持</span><span>Q 8.0</span><span class="trace-arrow">→</span><a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838"><code>S-003</code></a></div></div></article><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-004</strong><p class="trace-claim-text">须单独监测资深高技能客服的质量，不能假定所有岗位群体均受益。</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-004</code><span>Implementation Risk</span><span class="dir neg">负向效应</span><span class="trace-relation">主张关系: 支持</span><span>Q 8.0</span><span class="trace-arrow">→</span><a href="https://academic.oup.com/qje/article/140/2/889/7990658"><code>S-001</code></a></div></div></article></div><div id="chart-trace-zh" class="chart-mount" aria-label="主张-证据追溯"></div><p class="chart-interpretation"><strong>这意味着什么：</strong>每个重要主张都必须能追到 Evidence ID 和原始来源。</p></div></section><section id="full-04-action" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>04 适用范围与教学行动</h2><p class="full-chapter-lead">把可外推范围、护栏和教学动作连接到具体证据。</p></header><div class="full-chapter-body"><p><strong>目标人群：</strong>企业客服人员，按资历和基线技能分层。</p>
+</div><div class="scope-grid"><article class="scope-card"><h3>研究问题</h3><p class="scope-text">企业客服团队是否应引入生成式 AI 助手？</p></article><article class="scope-card"><h3>干预方案</h3><p class="scope-text">政策名称：人工监督的客服助手；政策类型：机构改革；作用机制：已批准知识库 + 坐席复核</p></article><article class="scope-card"><h3>比较条件</h3><p class="scope-text">不提供生成式 AI 建议的现有客服流程。</p></article><article class="scope-card"><h3>结果构念</h3><p class="scope-text">主要结果：政策有效性、实施风险；次要结果：成本效益、公平性、可行性</p></article><article class="scope-card"><h3>研究范围</h3><p class="scope-text">时间范围：2023–2026；Evidence Types：准实验、随机对照试验</p></article><article class="scope-card"><h3>决策成功条件</h3><p class="scope-text">提高每付薪工时的有效解决量，同时维持服务质量、隐私和员工自主判断。</p></article></div></div></section><section id="full-02-evidence" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>02 关键证据与结果分离</h2><p class="full-chapter-lead">把过程产出、真实效果、持续性与风险放在同一证据地图中，但不混为一谈。</p></header><div class="full-chapter-body"><div class="retrieval-grid"><article><h3>纳入标准</h3><ul></ul></article><article><h3>排除标准</h3><ul></ul></article></div><div class="retrieval-coverage"><h3>证据来源覆盖</h3><p><code>S-001</code> <code>S-002</code> <code>S-003</code></p><p class="retrieval-note">当前报告只展示 result 中真实存在的检索与来源信息；没有流程计数时不伪造 PRISMA / funnel 数字。</p></div><div class="outcome-separation" data-visual="outcome-separation"><div class="visual-heading"><h3>结果分离 · 过程产出 ≠ 政策效果</h3><p>将不同结果类型分开裁决，避免把试点期产出或效率提升直接等同于真实政策效果。</p><p class="semantic-note">效应方向来自 evidence.effect_direction；“支持某个主张”不等于“结果是正向”。</p></div><div class="outcome-groups"><article class="outcome-group outcome-learning"><h3>效果 / 目标结果</h3><ul><li><strong>政策有效性</strong><span class="outcome-states"><span class="dir pos">正向效应 2</span></span></li></ul></article><article class="outcome-group outcome-risk"><h3>风险 / 实施风险</h3><ul><li><strong>实施风险</strong><span class="outcome-states"><span class="dir neg">负向效应 2</span></span></li></ul></article></div></div><div class='table-wrap outcome-table'><table class='data-table'><thead><tr><th>结果类型</th><th>正向效应</th><th>负向效应</th><th>零效应</th><th>证据</th></tr></thead><tbody><tr><td><strong>政策有效性</strong><span class='raw-tag' title='原始标识'>policy_effectiveness</span></td><td class='num'>2</td><td class='num'>0</td><td class='num'>0</td><td><code>E-001</code> <code>E-002</code> </td></tr><tr><td><strong>实施风险</strong><span class='raw-tag' title='原始标识'>implementation_risk</span></td><td class='num'>0</td><td class='num'>2</td><td class='num'>0</td><td><code>E-003</code> <code>E-004</code> </td></tr></tbody></table></div><div class='matrix-controls'><div class='matrix-tools'><input id='matrix-search-full-zh' type='search' placeholder='搜索证据…' aria-label='筛选 / 搜索证据'><select id='matrix-direction-full-zh' aria-label='按效应方向筛选'><option value=''>全部效应</option><option value='positive'>正向效应</option><option value='negative'>负向效应</option><option value='null'>零效应</option></select><select id='matrix-outcome-full-zh' aria-label='按结果类型筛选'><option value=''>全部结果</option><option value='implementation_risk'>实施风险</option><option value='policy_effectiveness'>政策有效性</option></select></div></div><div class='table-wrap matrix-wrap'><table id='evidence-matrix-full-zh' class='data-table evidence-matrix'><thead><tr><th>ID</th><th>结果</th><th>效应</th><th>质量</th><th>主张</th><th>来源</th></tr></thead><tbody><tr data-effect="positive" data-direction="positive" data-outcome="policy_effectiveness" data-search="e-001 st-001 sample-support-all generative ai at work 在有边界的客服场景中，ai 辅助可缩短会话处理时间；吞吐量是另一项指标。 客服人员   positive s-001"><td><code>E-001</code></td><td><strong>政策有效性</strong></td><td><span class="dir pos">正向效应</span><span class="relation-note">与主张关系: 支持</span></td><td><div class="quality-cell"><span class="num">9</span><span class="quality-meter" aria-hidden="true"><i style="width:90%"></i></span></div></td><td class="claim-cell"><p>在有边界的客服场景中，AI 辅助可缩短会话处理时间；吞吐量是另一项指标。</p><details class="matrix-row-detail"><summary>查看完整证据</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>研究 ID</dt><dd>ST-001</dd></div><div class="evidence-detail-row"><dt>样本 ID</dt><dd>SAMPLE-support-all</dd></div><div class="evidence-detail-row"><dt>研究标题</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>来源标题</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>年份</dt><dd>2025</dd></div><div class="evidence-detail-row"><dt>研究设计</dt><dd>准实验</dd></div><div class="evidence-detail-row"><dt>研究人群</dt><dd>客服人员</dd></div><div class="evidence-detail-row"><dt>样本量</dt><dd>5172</dd></div><div class="evidence-detail-row"><dt>结果测量</dt><dd>会话处理时间；另列吞吐量摘要：每小时解决问题数约增加 15%。</dd></div><div class="evidence-detail-row"><dt>效应方向</dt><dd>正向效应</dd></div><div class="evidence-detail-row"><dt>与主张关系</dt><dd>支持</dd></div><div class="evidence-detail-row"><dt>局限</dt><dd>单一企业、单一工具、非随机分批上线；因果解释依赖识别假设。</dd></div><div class="evidence-detail-row"><dt>质量维度</dt><dd>D1 研究设计=1 · D2 样本质量=2 · D3 测量效度=2 · D4 时间强度=2 · D5 直接性=2</dd></div><div class="evidence-detail-row"><dt>质量分</dt><dd>9.0</dd></div><div class="evidence-detail-row"><dt>证据状态</dt><dd>已支持</dd></div><div class="evidence-detail-row"><dt>完整主张</dt><dd>在有边界的客服场景中，AI 辅助可缩短会话处理时间；吞吐量是另一项指标。</dd></div><div class="evidence-detail-row"><dt>来源位置</dt><dd>https://academic.oup.com/qje/article/140/2/889/7990658</dd></div><div class="evidence-detail-row"><dt>可验证链接</dt><dd><a href="https://academic.oup.com/qje/article/140/2/889/7990658">https://academic.oup.com/qje/article/140/2/889/7990658</a></dd></div></dl></details></td><td><a class="source-link" href="https://academic.oup.com/qje/article/140/2/889/7990658" title="Generative AI at Work"><code>S-001</code></a></td></tr><tr data-effect="positive" data-direction="positive" data-outcome="policy_effectiveness" data-search="e-002 st-002 sample-writing experimental evidence on the productivity effects of generative artificial intelligence chatgpt 缩短了短篇职业写作任务用时；对客服而言属于间接证据。 受过大学教育的职场人士   positive s-002"><td><code>E-002</code></td><td><strong>政策有效性</strong></td><td><span class="dir pos">正向效应</span><span class="relation-note">与主张关系: 支持</span></td><td><div class="quality-cell"><span class="num">7</span><span class="quality-meter" aria-hidden="true"><i style="width:70%"></i></span></div></td><td class="claim-cell"><p>ChatGPT 缩短了短篇职业写作任务用时；对客服而言属于间接证据。</p><details class="matrix-row-detail"><summary>查看完整证据</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>研究 ID</dt><dd>ST-002</dd></div><div class="evidence-detail-row"><dt>样本 ID</dt><dd>SAMPLE-writing</dd></div><div class="evidence-detail-row"><dt>研究标题</dt><dd>Experimental evidence on the productivity effects of generative artificial intelligence</dd></div><div class="evidence-detail-row"><dt>来源标题</dt><dd>Experimental evidence on the productivity effects of generative artificial intelligence</dd></div><div class="evidence-detail-row"><dt>年份</dt><dd>2023</dd></div><div class="evidence-detail-row"><dt>研究设计</dt><dd>随机对照试验</dd></div><div class="evidence-detail-row"><dt>研究人群</dt><dd>受过大学教育的职场人士</dd></div><div class="evidence-detail-row"><dt>样本量</dt><dd>453</dd></div><div class="evidence-detail-row"><dt>结果测量</dt><dd>自报任务用时约减少 40%；评定写作质量约提高 18% 是另一项指标。</dd></div><div class="evidence-detail-row"><dt>效应方向</dt><dd>正向效应</dd></div><div class="evidence-detail-row"><dt>与主张关系</dt><dd>支持</dd></div><div class="evidence-detail-row"><dt>局限</dt><dd>间接证据，不可直接推广：短时激励写作任务不要求精确事实或客户特定情境。</dd></div><div class="evidence-detail-row"><dt>质量维度</dt><dd>D1 研究设计=2 · D2 样本质量=2 · D3 测量效度=1 · D4 时间强度=1 · D5 直接性=1</dd></div><div class="evidence-detail-row"><dt>质量分</dt><dd>7.0</dd></div><div class="evidence-detail-row"><dt>证据状态</dt><dd>已支持</dd></div><div class="evidence-detail-row"><dt>完整主张</dt><dd>ChatGPT 缩短了短篇职业写作任务用时；对客服而言属于间接证据。</dd></div><div class="evidence-detail-row"><dt>来源位置</dt><dd>https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</dd></div><div class="evidence-detail-row"><dt>可验证链接</dt><dd><a href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf">https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</a></dd></div></dl></details></td><td><a class="source-link" href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf" title="Experimental evidence on the productivity effects of generative artificial intelligence"><code>S-002</code></a></td></tr><tr data-effect="negative" data-direction="negative" data-outcome="implementation_risk" data-search="e-003 st-003 sample-consulting-outside navigating the jagged technological frontier: field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality 在超出能力边界的任务中 ai 可能降低正确率；咨询实验对客服属于间接证据。 参与超出能力边界实验的 bcg 顾问   negative s-003"><td><code>E-003</code></td><td><strong>实施风险</strong></td><td><span class="dir neg">负向效应</span><span class="relation-note">与主张关系: 支持</span></td><td><div class="quality-cell"><span class="num">8</span><span class="quality-meter" aria-hidden="true"><i style="width:80%"></i></span></div></td><td class="claim-cell"><p>在超出能力边界的任务中 AI 可能降低正确率；咨询实验对客服属于间接证据。</p><details class="matrix-row-detail"><summary>查看完整证据</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>研究 ID</dt><dd>ST-003</dd></div><div class="evidence-detail-row"><dt>样本 ID</dt><dd>SAMPLE-consulting-outside</dd></div><div class="evidence-detail-row"><dt>研究标题</dt><dd>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality</dd></div><div class="evidence-detail-row"><dt>来源标题</dt><dd>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality</dd></div><div class="evidence-detail-row"><dt>年份</dt><dd>2026</dd></div><div class="evidence-detail-row"><dt>研究设计</dt><dd>随机对照试验</dd></div><div class="evidence-detail-row"><dt>研究人群</dt><dd>参与超出能力边界实验的 BCG 顾问</dd></div><div class="evidence-detail-row"><dt>样本量</dt><dd>373</dd></div><div class="evidence-detail-row"><dt>结果测量</dt><dd>商业建议正确率：合并 AI 组约低 19 个百分点；表 7 为 373 人，不是 758 人。</dd></div><div class="evidence-detail-row"><dt>效应方向</dt><dd>负向效应</dd></div><div class="evidence-detail-row"><dt>与主张关系</dt><dd>支持</dd></div><div class="evidence-detail-row"><dt>局限</dt><dd>间接证据，不可直接推广：咨询顾问处理实验商业案例，并非真实客服工单。</dd></div><div class="evidence-detail-row"><dt>质量维度</dt><dd>D1 研究设计=2 · D2 样本质量=2 · D3 测量效度=2 · D4 时间强度=1 · D5 直接性=1</dd></div><div class="evidence-detail-row"><dt>质量分</dt><dd>8.0</dd></div><div class="evidence-detail-row"><dt>证据状态</dt><dd>已支持</dd></div><div class="evidence-detail-row"><dt>完整主张</dt><dd>在超出能力边界的任务中 AI 可能降低正确率；咨询实验对客服属于间接证据。</dd></div><div class="evidence-detail-row"><dt>来源位置</dt><dd>https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</dd></div><div class="evidence-detail-row"><dt>可验证链接</dt><dd><a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838">https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</a></dd></div></dl></details></td><td><a class="source-link" href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838" title="Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality"><code>S-003</code></a></td></tr><tr data-effect="negative" data-direction="negative" data-outcome="implementation_risk" data-search="e-004 st-001 sample-support-all generative ai at work 须单独监测资深高技能客服的质量，不能假定所有岗位群体均受益。 资深高技能客服人员   negative s-001"><td><code>E-004</code></td><td><strong>实施风险</strong></td><td><span class="dir neg">负向效应</span><span class="relation-note">与主张关系: 支持</span></td><td><div class="quality-cell"><span class="num">8</span><span class="quality-meter" aria-hidden="true"><i style="width:80%"></i></span></div></td><td class="claim-cell"><p>须单独监测资深高技能客服的质量，不能假定所有岗位群体均受益。</p><details class="matrix-row-detail"><summary>查看完整证据</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>研究 ID</dt><dd>ST-001</dd></div><div class="evidence-detail-row"><dt>样本 ID</dt><dd>SAMPLE-support-all</dd></div><div class="evidence-detail-row"><dt>研究标题</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>来源标题</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>年份</dt><dd>2025</dd></div><div class="evidence-detail-row"><dt>研究设计</dt><dd>准实验</dd></div><div class="evidence-detail-row"><dt>研究人群</dt><dd>资深高技能客服人员</dd></div><div class="evidence-detail-row"><dt>结果测量</dt><dd>资历最深、技能最高的客服群体出现小幅质量下降。</dd></div><div class="evidence-detail-row"><dt>效应方向</dt><dd>负向效应</dd></div><div class="evidence-detail-row"><dt>与主张关系</dt><dd>支持</dd></div><div class="evidence-detail-row"><dt>局限</dt><dd>与 E-001 属于同一研究；未提取该亚组样本量，不是独立重复验证。</dd></div><div class="evidence-detail-row"><dt>质量维度</dt><dd>D1 研究设计=1 · D2 样本质量=1 · D3 测量效度=2 · D4 时间强度=2 · D5 直接性=2</dd></div><div class="evidence-detail-row"><dt>质量分</dt><dd>8.0</dd></div><div class="evidence-detail-row"><dt>证据状态</dt><dd>已支持</dd></div><div class="evidence-detail-row"><dt>完整主张</dt><dd>须单独监测资深高技能客服的质量，不能假定所有岗位群体均受益。</dd></div><div class="evidence-detail-row"><dt>来源位置</dt><dd>https://academic.oup.com/qje/article/140/2/889/7990658</dd></div><div class="evidence-detail-row"><dt>可验证链接</dt><dd><a href="https://academic.oup.com/qje/article/140/2/889/7990658">https://academic.oup.com/qje/article/140/2/889/7990658</a></dd></div></dl></details></td><td><a class="source-link" href="https://academic.oup.com/qje/article/140/2/889/7990658" title="Generative AI at Work"><code>S-001</code></a></td></tr></tbody></table></div></div></section><section id="full-03-quality" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>03 证据可信度、反证与方法审计</h2><p class="full-chapter-lead">检查证据为什么可信、哪里冲突，以及哪些结论必须降级。</p></header><div class="full-chapter-body"><section class="method-review-group"><header class="method-review-title"><h3>审查目标：overall</h3><span class="method-verdict">关注</span></header><div class="method-audit-grid"><article class="method-audit-item method-met"><div class="method-audit-head"><strong>证据等级</strong><span class="method-status">通过</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>因果识别</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>外部效度</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-missing"><div class="method-audit-head"><strong>成本证据</strong><span class="method-status">缺失</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>利益相关者代表</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>实施证据</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>公平性分析</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>效应量报告</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>不确定性量化</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-met"><div class="method-audit-head"><strong>对照与反事实清晰</strong><span class="method-status">通过</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>发表偏倚风险</strong><span class="method-status">部分</span></div><p></p></article><article class="method-audit-item method-met"><div class="method-audit-head"><strong>可推广性声明</strong><span class="method-status">通过</span></div><p></p></article></div><div class="method-guard-wrap"><strong>产出 vs 效果护栏：</strong><p class="method-guard">教学结果不适用；生产率并非学生学习效果主张。</p></div></section><p>无数据。</p><div class="evidence-tribunal" data-visual="evidence-tribunal-grid"><p class="tribunal-summary"><strong>决策：</strong>试点验证 · <strong>置信度：</strong>中</p><div class="tribunal-grid"><article class="tribunal-card supported"><header><span aria-hidden="true">✓</span><h3>可以主张</h3><span class="tribunal-count">4</span></header><ul><li><p>在有边界的客服场景中，AI 辅助可缩短会话处理时间；吞吐量是另一项指标。 — E-001</p><div class="tribunal-evidence-refs"><code>E-001</code></div></li><li><p>ChatGPT 缩短了短篇职业写作任务用时；对客服而言属于间接证据。 — E-002</p><div class="tribunal-evidence-refs"><code>E-002</code></div></li><li><p>在超出能力边界的任务中 AI 可能降低正确率；咨询实验对客服属于间接证据。 — E-003</p><div class="tribunal-evidence-refs"><code>E-003</code></div></li><li><p>须单独监测资深高技能客服的质量，不能假定所有岗位群体均受益。 — E-004</p><div class="tribunal-evidence-refs"><code>E-004</code></div></li></ul></article><article class="tribunal-card uncertain"><header><span aria-hidden="true">?</span><h3>尚不能主张</h3><span class="tribunal-count">1</span></header><ul><li><p>这组选编证据尚未确立本企业净成本、隐私事件率、亚组服务质量及持续表现 [无直接证据]。</p></li></ul></article><article class="tribunal-card contradicted"><header><span aria-hidden="true">×</span><h3>被反驳的主张</h3><span class="tribunal-count">0</span></header><ul><li>无数据。</li></ul></article><article class="tribunal-card missing"><header><span aria-hidden="true">…</span><h3>缺失证据</h3><span class="tribunal-count">1</span></header><ul><li><p>这组选编证据尚未确立本企业净成本、隐私事件率、亚组服务质量及持续表现 [无直接证据]。</p></li></ul></article></div><details class="supporting-visual"><summary>EvidenceFlow 协议</summary><svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="EvidenceFlow 协议流程"><title>EvidenceFlow 协议流程</title><desc>从问题框架、检索、抓取验证、证据抽取、反方质疑、方法审计、裁决到适用性与干预评价的完整流程。</desc><defs><marker id="arr-zh-full-workflow" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">EvidenceFlow 协议</text><rect x="24" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="56.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="56.0" y="126.0">问题框架</tspan></text><line x1="88" y1="126" x2="100" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="100" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="132.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="132.0" y="126.0">检索</tspan></text><line x1="164" y1="126" x2="176" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="176" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="208.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="208.0" y="126.0">抓取</tspan><tspan x="208.0" y="140.0">验证</tspan></text><line x1="240" y1="126" x2="252" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="252" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="284.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="284.0" y="126.0">证据抽取</tspan></text><line x1="316" y1="126" x2="328" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="328" y="100" width="64" height="52" rx="8" fill="#5E8A6A"/><text x="360.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="360.0" y="126.0">反方质疑</tspan></text><line x1="392" y1="126" x2="404" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="404" y="100" width="64" height="52" rx="8" fill="#5E8A6A"/><text x="436.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="436.0" y="126.0">方法</tspan><tspan x="436.0" y="140.0">审计</tspan></text><line x1="468" y1="126" x2="480" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="480" y="100" width="64" height="52" rx="8" fill="#C99A4A"/><text x="512.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="512.0" y="126.0">裁决</tspan></text><line x1="544" y1="126" x2="556" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="556" y="100" width="64" height="52" rx="8" fill="#C99A4A"/><text x="588.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="588.0" y="126.0">适用性</tspan></text><line x1="620" y1="126" x2="632" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-workflow)"/><rect x="632" y="100" width="64" height="52" rx="8" fill="#C99A4A"/><text x="664.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="664.0" y="126.0">干预</tspan><tspan x="664.0" y="140.0">评价</tspan></text></svg></details><details class="supporting-visual"><summary>裁决信息图</summary><svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="证据裁决信息图"><title>证据裁决信息图</title><desc>可以主张与不可主张的证据 ID 与建议决策徽章；完整主张文本见下方裁决卡片。</desc><defs><marker id="arr-zh-full-tribunal" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">证据裁决</text><text x="24" y="180" font-size="13" font-weight="700" fill="#3A3833">冲突来源</text><text x="24" y="202" font-size="11" fill="#8A867E">详见下方裁决卡片</text><rect x="24" y="216" width="180" height="28" rx="14" fill="#C99A4A"/><text x="114" y="235" text-anchor="middle" font-size="13" font-weight="700" fill="#fff">试点验证</text><text x="250" y="80" font-size="13" font-weight="700" fill="#5E8A6A">可以主张 (4)</text><circle cx="254" cy="100" r="3" fill="#5E8A6A"/><text x="266" y="105" font-size="11" fill="#3A3833">E-001</text><circle cx="254" cy="122" r="3" fill="#5E8A6A"/><text x="266" y="127" font-size="11" fill="#3A3833">E-002</text><circle cx="254" cy="144" r="3" fill="#5E8A6A"/><text x="266" y="149" font-size="11" fill="#3A3833">E-003</text><circle cx="254" cy="166" r="3" fill="#5E8A6A"/><text x="266" y="171" font-size="11" fill="#3A3833">E-004</text><text x="490" y="80" font-size="13" font-weight="700" fill="#A85B53">不可主张 (1)</text><circle cx="494" cy="100" r="3" fill="#A85B53"/><text x="506" y="105" font-size="11" fill="#3A3833">（无）</text></svg></details></div><div class="trace-chain" data-visual="trace-chain"><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-001</strong><p class="trace-claim-text">在有边界的客服场景中，AI 辅助可缩短会话处理时间；吞吐量是另一项指标。</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-001</code><span>政策有效性</span><span class="dir pos">正向效应</span><span class="trace-relation">主张关系: 支持</span><span>Q 9.0</span><span class="trace-arrow">→</span><a href="https://academic.oup.com/qje/article/140/2/889/7990658"><code>S-001</code></a></div></div></article><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-002</strong><p class="trace-claim-text">ChatGPT 缩短了短篇职业写作任务用时；对客服而言属于间接证据。</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-002</code><span>政策有效性</span><span class="dir pos">正向效应</span><span class="trace-relation">主张关系: 支持</span><span>Q 7.0</span><span class="trace-arrow">→</span><a href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf"><code>S-002</code></a></div></div></article><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-003</strong><p class="trace-claim-text">在超出能力边界的任务中 AI 可能降低正确率；咨询实验对客服属于间接证据。</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-003</code><span>实施风险</span><span class="dir neg">负向效应</span><span class="trace-relation">主张关系: 支持</span><span>Q 8.0</span><span class="trace-arrow">→</span><a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838"><code>S-003</code></a></div></div></article><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-004</strong><p class="trace-claim-text">须单独监测资深高技能客服的质量，不能假定所有岗位群体均受益。</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-004</code><span>实施风险</span><span class="dir neg">负向效应</span><span class="trace-relation">主张关系: 支持</span><span>Q 8.0</span><span class="trace-arrow">→</span><a href="https://academic.oup.com/qje/article/140/2/889/7990658"><code>S-001</code></a></div></div></article></div><div id="chart-trace-zh" class="chart-mount" aria-label="主张-证据追溯"></div><p class="chart-interpretation"><strong>这意味着什么：</strong>每个重要主张都必须能追到 Evidence ID 和原始来源。</p></div></section><section id="full-04-action" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>04 适用范围与政策行动</h2><p class="full-chapter-lead">把可外推范围、护栏和政策动作连接到具体证据。</p></header><div class="full-chapter-body"><p><strong>目标人群：</strong>企业客服人员，按资历和基线技能分层。</p>
 <p><strong>目标情境：</strong>使用经审核知识库、由人工监督的客服流程。</p>
 <p><strong>适用条件：</strong></p><ul><li>从低风险且知识库覆盖的队列开始；客服审查并编辑每条建议回复。</li><li>按资历和基线技能分层；保留专家判断并记录复核工作量。</li><li>试点前落实客户数据最小化与脱敏、访问和留存限制，并审查供应商数据使用条款。</li><li>陌生、模糊、敏感或高风险请求交由合格人员处理；禁止自动作出承诺。</li></ul><div class="boundary-block"><h3>不可外推的结论</h3><ul><li>声称普遍收益或自动部署安全超出证据边界：没有纳入研究测量隐私事件、本地净成本或亚组服务质量。</li></ul></div><div class="evidence-to-action" data-visual="evidence-to-action"><article class="action-node action-evidence"><span>证据</span><p class="action-node-text">在有边界的客服场景中，AI 辅助可缩短会话处理时间；吞吐量是另一项指标。 — E-001</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-applicability"><span>适用性</span><p class="action-node-text">从低风险且知识库覆盖的队列开始；客服审查并编辑每条建议回复。; 按资历和基线技能分层；保留专家判断并记录复核工作量。; 试点前落实客户数据最小化与脱敏、访问和留存限制，并审查供应商数据使用条款。; 陌生、模糊、敏感或高风险请求交由合格人员处理；禁止自动作出承诺。</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-decision"><span>决策</span><p class="action-node-text">试点验证</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-guardrails"><span>护栏</span><p class="action-node-text">从低风险且知识库覆盖的队列开始；客服审查并编辑每条建议回复。</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-stop"><span>停止条件</span><p class="action-node-text">一旦核实隐私泄漏、严重不安全建议或越权客户承诺，立即暂停，调查后再决定是否恢复。; 若盲评质量越过预先约定的非劣界值（包括各资历分层），暂停扩展。</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-evaluation"><span>评价</span><p class="action-node-text">仅在质量非劣、每付薪工时有效解决量提高、净成本可接受且无未解决严重安全事件时考虑扩展；阈值须在本地预先约定。</p></article></div><p><strong>目标人群：</strong>企业客服人员，按资历和基线技能分层。 · <strong>试点时长：</strong>建议：两周基线和六周试点；尚未执行。</p>
-<p><strong>AI 使用规则：</strong>从低风险且知识库覆盖的队列开始；客服审查并编辑每条建议回复。</p>
+<p><strong>干预使用规则：</strong>从低风险且知识库覆盖的队列开始；客服审查并编辑每条建议回复。</p>
 <h3>停止条件</h3><ul><li>一旦核实隐私泄漏、严重不安全建议或越权客户承诺，立即暂停，调查后再决定是否恢复。</li><li>若盲评质量越过预先约定的非劣界值（包括各资历分层），暂停扩展。</li></ul>
-<h3>干预时间线信息图</h3>
-<svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="教学干预时间线"><title>教学干预时间线</title><desc>各试点阶段的短名称与活动数量；完整 AI 使用规则见阶段说明块。</desc><defs><marker id="arr-zh-full-intervention" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">教学干预时间线</text><rect x="24" y="100" width="160" height="96" rx="8" fill="#B8694A"/><text x="104.0" y="148.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="104.0" y="148.0">Phase 1</tspan></text></svg></div></section><section id="full-05-evaluation" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>05 试点设计、评估与停止条件</h2><p class="full-chapter-lead">用独立学习结果验证试点，并预先写清停止条件。</p></header><div class="full-chapter-body"><p><strong>研究问题：</strong>企业客服团队是否应引入生成式 AI 助手？</p>
+<h3>干预方案时间线信息图</h3>
+<svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="干预方案时间线"><title>干预方案时间线</title><desc>各试点阶段的短名称与活动数量；完整干预使用规则见阶段说明块。</desc><defs><marker id="arr-zh-full-intervention" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">干预方案时间线</text><rect x="24" y="100" width="160" height="96" rx="8" fill="#B8694A"/><text x="104.0" y="148.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="104.0" y="148.0">Phase 1</tspan></text></svg></div></section><section id="full-05-evaluation" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>05 试点设计、评估与停止条件</h2><p class="full-chapter-lead">用独立效果测量验证试点，并预先写清停止条件。</p></header><div class="full-chapter-body"><p><strong>研究问题：</strong>企业客服团队是否应引入生成式 AI 助手？</p>
 <p><strong>基线：</strong>分配前记录解决率、付薪工时、重复联系、盲评质量及复核成本。</p>
-<p><strong>后测：</strong>试点结束时评估相同结果；独立审计隐私与不安全承诺。</p>
+<p><strong>试点结束：</strong>试点结束时评估相同结果；独立审计隐私与不安全承诺。</p>
 <p><strong>成功阈值：</strong>仅在质量非劣、每付薪工时有效解决量提高、净成本可接受且无未解决严重安全事件时考虑扩展；阈值须在本地预先约定。</p>
 <p><strong>分析计划：</strong>入组前预注册意向治疗比较、团队聚类不确定性及亚组估计。根据本地基线变异和业务目标确定样本量与非劣界值。</p>
 <h3>评价设计信息图</h3>
-<svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="评价设计流程"><title>评价设计流程</title><desc>基线、后测、保持测试与迁移测试的评价流程；完整指标与分析计划见评估章节。</desc><defs><marker id="arr-zh-full-evaluation" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">评价设计流程</text><rect x="30" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="105.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="105.0" y="138.0">基线</tspan></text><text x="105.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">前测</text><line x1="180" y1="138" x2="192" y2="138" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-evaluation)"/><rect x="192" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="267.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="267.0" y="138.0">后测</tspan></text><text x="267.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">后测</text><line x1="342" y1="138" x2="354" y2="138" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-evaluation)"/><rect x="354" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="429.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="429.0" y="138.0">保持</tspan></text><text x="429.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">保持测试</text><line x1="504" y1="138" x2="516" y2="138" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-evaluation)"/><rect x="516" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="591.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="591.0" y="138.0">迁移</tspan></text><text x="591.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">迁移测试（无 AI）</text></svg><div class="visual-suppressed"><strong>基准图已抑制</strong><p>result.json 未携带 benchmark.baselines，本图不绘制；基准表现见独立基准报告。</p></div></div></section><section id="full-06-sources" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>06 来源、溯源与附录</h2><p class="full-chapter-lead">保留原始来源、URL、证据 ID 和获取信息，确保可回查。</p></header><div class="full-chapter-body"><h3>来源列表</h3><div class='table-wrap'><table class='data-table source-table'><thead><tr><th>ID</th><th>标题</th><th>年份</th><th>权威级别</th><th>可验证位置</th></tr></thead><tbody><tr><td><code>S-001</code></td><td class='cell-main source-title-cell'>Generative AI at Work<details class="source-expander detail-expander"><summary>查看来源与溯源</summary><dl class="source-detail-grid"><div class="source-detail-row"><dt>原文标题</dt><dd>Generative AI at Work</dd></div><div class="source-detail-row"><dt>年份</dt><dd>2025</dd></div><div class="source-detail-row"><dt>权威级别</dt><dd>T1 DOI 可验证论文</dd></div><div class="source-detail-row"><dt>获取方式</dt><dd>builtin</dd></div><div class="source-detail-row"><dt>获取状态</dt><dd>FETCH_VALID</dd></div><div class="source-detail-row"><dt>可验证链接</dt><dd><a href="https://academic.oup.com/qje/article/140/2/889/7990658">https://academic.oup.com/qje/article/140/2/889/7990658</a></dd></div></dl></details></td><td>2025</td><td>T1 DOI 可验证论文</td><td class='cell-main'><a href='https://academic.oup.com/qje/article/140/2/889/7990658'>https://academic.oup.com/qje/article/140/2/889/7990658</a></td></tr><tr><td><code>S-002</code></td><td class='cell-main source-title-cell'>Experimental evidence on the productivity effects of generative artificial intelligence<details class="source-expander detail-expander"><summary>查看来源与溯源</summary><dl class="source-detail-grid"><div class="source-detail-row"><dt>原文标题</dt><dd>Experimental evidence on the productivity effects of generative artificial intelligence</dd></div><div class="source-detail-row"><dt>年份</dt><dd>2023</dd></div><div class="source-detail-row"><dt>权威级别</dt><dd>T1 DOI 可验证论文</dd></div><div class="source-detail-row"><dt>获取方式</dt><dd>builtin</dd></div><div class="source-detail-row"><dt>获取状态</dt><dd>FETCH_VALID</dd></div><div class="source-detail-row"><dt>可验证链接</dt><dd><a href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf">https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</a></dd></div></dl></details></td><td>2023</td><td>T1 DOI 可验证论文</td><td class='cell-main'><a href='https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf'>https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</a></td></tr><tr><td><code>S-003</code></td><td class='cell-main source-title-cell'>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality<details class="source-expander detail-expander"><summary>查看来源与溯源</summary><dl class="source-detail-grid"><div class="source-detail-row"><dt>原文标题</dt><dd>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality</dd></div><div class="source-detail-row"><dt>年份</dt><dd>2026</dd></div><div class="source-detail-row"><dt>权威级别</dt><dd>T1 DOI 可验证论文</dd></div><div class="source-detail-row"><dt>获取方式</dt><dd>builtin</dd></div><div class="source-detail-row"><dt>获取状态</dt><dd>FETCH_VALID</dd></div><div class="source-detail-row"><dt>可验证链接</dt><dd><a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838">https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</a></dd></div></dl></details></td><td>2026</td><td>T1 DOI 可验证论文</td><td class='cell-main'><a href='https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838'>https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</a></td></tr></tbody></table></div><h3>Fetch 溯源</h3><div class='table-wrap'><table class='data-table'><thead><tr><th>来源</th><th>Fetch 方式</th><th>状态</th><th>降级</th><th>时间</th></tr></thead><tbody><tr><td><code>S-001</code></td><td>builtin</td><td>FETCH_VALID</td><td></td><td></td></tr><tr><td><code>S-002</code></td><td>builtin</td><td>FETCH_VALID</td><td></td><td></td></tr><tr><td><code>S-003</code></td><td>builtin</td><td>FETCH_VALID</td><td></td><td></td></tr></tbody></table></div></div></section></main></div>
+<svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="评价设计流程"><title>评价设计流程</title><desc>基线、试点结束、持续监测与扩展评估的评价流程；完整指标与分析计划见评估章节。</desc><defs><marker id="arr-zh-full-evaluation" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">评价设计流程</text><rect x="30" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="105.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="105.0" y="138.0">基线</tspan></text><text x="105.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">前测</text><line x1="180" y1="138" x2="192" y2="138" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-evaluation)"/><rect x="192" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="267.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="267.0" y="138.0">后测</tspan></text><text x="267.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">后测</text><line x1="342" y1="138" x2="354" y2="138" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-evaluation)"/><rect x="354" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="429.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="429.0" y="138.0">保持</tspan></text><text x="429.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">持续监测</text><line x1="504" y1="138" x2="516" y2="138" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-zh-full-evaluation)"/><rect x="516" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="591.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="591.0" y="138.0">迁移</tspan></text><text x="591.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">扩展评估</text></svg><div class="visual-suppressed"><strong>基准图已抑制</strong><p>result.json 未携带 benchmark.baselines，本图不绘制；基准表现见独立基准报告。</p></div></div></section><section id="full-06-sources" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>06 来源、溯源与附录</h2><p class="full-chapter-lead">保留原始来源、URL、证据 ID 和获取信息，确保可回查。</p></header><div class="full-chapter-body"><h3>来源列表</h3><div class='table-wrap'><table class='data-table source-table'><thead><tr><th>ID</th><th>标题</th><th>年份</th><th>权威级别</th><th>可验证位置</th></tr></thead><tbody><tr><td><code>S-001</code></td><td class='cell-main source-title-cell'>Generative AI at Work<details class="source-expander detail-expander"><summary>查看来源与溯源</summary><dl class="source-detail-grid"><div class="source-detail-row"><dt>原文标题</dt><dd>Generative AI at Work</dd></div><div class="source-detail-row"><dt>年份</dt><dd>2025</dd></div><div class="source-detail-row"><dt>权威级别</dt><dd>T1 DOI 可验证论文</dd></div><div class="source-detail-row"><dt>获取方式</dt><dd>builtin</dd></div><div class="source-detail-row"><dt>获取状态</dt><dd>FETCH_VALID</dd></div><div class="source-detail-row"><dt>可验证链接</dt><dd><a href="https://academic.oup.com/qje/article/140/2/889/7990658">https://academic.oup.com/qje/article/140/2/889/7990658</a></dd></div></dl></details></td><td>2025</td><td>T1 DOI 可验证论文</td><td class='cell-main'><a href='https://academic.oup.com/qje/article/140/2/889/7990658'>https://academic.oup.com/qje/article/140/2/889/7990658</a></td></tr><tr><td><code>S-002</code></td><td class='cell-main source-title-cell'>Experimental evidence on the productivity effects of generative artificial intelligence<details class="source-expander detail-expander"><summary>查看来源与溯源</summary><dl class="source-detail-grid"><div class="source-detail-row"><dt>原文标题</dt><dd>Experimental evidence on the productivity effects of generative artificial intelligence</dd></div><div class="source-detail-row"><dt>年份</dt><dd>2023</dd></div><div class="source-detail-row"><dt>权威级别</dt><dd>T1 DOI 可验证论文</dd></div><div class="source-detail-row"><dt>获取方式</dt><dd>builtin</dd></div><div class="source-detail-row"><dt>获取状态</dt><dd>FETCH_VALID</dd></div><div class="source-detail-row"><dt>可验证链接</dt><dd><a href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf">https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</a></dd></div></dl></details></td><td>2023</td><td>T1 DOI 可验证论文</td><td class='cell-main'><a href='https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf'>https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</a></td></tr><tr><td><code>S-003</code></td><td class='cell-main source-title-cell'>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality<details class="source-expander detail-expander"><summary>查看来源与溯源</summary><dl class="source-detail-grid"><div class="source-detail-row"><dt>原文标题</dt><dd>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality</dd></div><div class="source-detail-row"><dt>年份</dt><dd>2026</dd></div><div class="source-detail-row"><dt>权威级别</dt><dd>T1 DOI 可验证论文</dd></div><div class="source-detail-row"><dt>获取方式</dt><dd>builtin</dd></div><div class="source-detail-row"><dt>获取状态</dt><dd>FETCH_VALID</dd></div><div class="source-detail-row"><dt>可验证链接</dt><dd><a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838">https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</a></dd></div></dl></details></td><td>2026</td><td>T1 DOI 可验证论文</td><td class='cell-main'><a href='https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838'>https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</a></td></tr></tbody></table></div><h3>Fetch 溯源</h3><div class='table-wrap'><table class='data-table'><thead><tr><th>来源</th><th>Fetch 方式</th><th>状态</th><th>降级</th><th>时间</th></tr></thead><tbody><tr><td><code>S-001</code></td><td>builtin</td><td>FETCH_VALID</td><td></td><td></td></tr><tr><td><code>S-002</code></td><td>builtin</td><td>FETCH_VALID</td><td></td><td></td></tr><tr><td><code>S-003</code></td><td>builtin</td><td>FETCH_VALID</td><td></td><td></td></tr></tbody></table></div></div></section></main></div>
 </div>
 <footer class="report-footer"><p>EduEvidence 证据报告 · Schema PASS · Claim Binding PASS · Numeric Consistency PASS · Bilingual Structure PASS · 语言人话化 PASS · 无伪精度 PASS · Lieflat 数据溯源 PASS · 坐标轴无失真 NOT_CHECKED · 色盲安全 NOT_CHECKED · 单文件离线可打开 · 数据源：result.json</p></footer>
 </div>
@@ -2173,7 +2173,7 @@ select:focus-visible,
 <button type="button" class="report-view-btn" data-report-view="full" aria-pressed="false">Full Report</button>
 </nav>
 </header>
-<div class="report-page report-page-brief" data-report-page="brief"><nav class="brief-navigation" aria-label="Brief navigation"><a href="#brief-decision-en"><span>01</span>Decision first</a><a href="#brief-outcomes-en"><span>02</span>Task performance ≠ learning</a><a href="#brief-tribunal-en"><span>03</span>Evidence tribunal</a><a href="#brief-action-en"><span>04</span>Evidence to action</a><a href="#brief-sources-en"><span>05</span>Key sources</a></nav><div class="brief-reading-content"><section class="brief-block brief-decision" id="brief-decision-en"><header class="brief-block-header"><h2>Decision first</h2><p>What to do, how confident we are, and the most important evidence boundary.</p></header><div class="brief-block-body">
+<div class="report-page report-page-brief" data-report-page="brief"><nav class="brief-navigation" aria-label="Brief navigation"><a href="#brief-decision-en"><span>01</span>Decision first</a><a href="#brief-outcomes-en"><span>02</span>Process output ≠ policy effect</a><a href="#brief-tribunal-en"><span>03</span>Evidence tribunal</a><a href="#brief-action-en"><span>04</span>Evidence to action</a><a href="#brief-sources-en"><span>05</span>Key sources</a></nav><div class="brief-reading-content"><section class="brief-block brief-decision" id="brief-decision-en"><header class="brief-block-header"><h2>Decision first</h2><p>What to do, how confident we are, and the most important evidence boundary.</p></header><div class="brief-block-body">
 <div class="decision-hero pilot" data-visual="decision-hero">
   <div class="hero-decision">
     <span class="eyebrow">Recommended decision</span>
@@ -2188,7 +2188,7 @@ select:focus-visible,
     <article class="hero-insight next"><span>Next action</span><p class="hero-insight-text">Run a supervised pilot on approved knowledge bases with human review on every reply, and track escalation and correction rates.</p></article>
   </div>
   <p class="hero-provenance"><span>Evidence / sources</span> · 4 / 3</p>
-</div></div></section><section class="brief-block brief-outcomes" id="brief-outcomes-en"><header class="brief-block-header"><h2>Task performance ≠ learning</h2><p>Only informative outcome separation; positive, negative and null effects use effect_direction.</p></header><div class="brief-block-body"><div class="outcome-separation" data-visual="outcome-separation"><div class="visual-heading"><h3>Outcome Separation · Task performance ≠ learning</h3><p>Outcomes are adjudicated separately so faster training performance is not silently treated as evidence of learning.</p><p class="semantic-note">Effect direction comes from evidence.effect_direction; supporting a claim does not imply a positive outcome.</p></div><div class="outcome-groups"><article class="outcome-group outcome-other"><h3>Other outcomes</h3><ul><li><strong>Policy Effectiveness</strong><span class="outcome-states"><span class="dir pos">Positive effect 2</span></span></li><li><strong>Implementation Risk</strong><span class="outcome-states"><span class="dir neg">Negative effect 2</span></span></li></ul></article></div></div></div></section><section class="brief-block brief-tribunal" id="brief-tribunal-en"><header class="brief-block-header"><h2>Evidence tribunal</h2><p>Supported, uncertain, contradicted and missing evidence stay separated instead of flattened into long prose.</p></header><div class="brief-block-body"><div class="evidence-tribunal" data-visual="evidence-tribunal-grid"><p class="tribunal-summary"><strong>Decision: </strong>Pilot · <strong>Confidence: </strong>Moderate</p><div class="tribunal-grid"><article class="tribunal-card supported"><header><span aria-hidden="true">✓</span><h3>Can claim</h3><span class="tribunal-count">4</span></header><ul><li><p>AI assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome. — E-001</p><div class="tribunal-evidence-refs"><code>E-001</code></div></li><li><p>ChatGPT reduced time on short professional writing tasks; this is indirect evidence for customer support. — E-002</p><div class="tribunal-evidence-refs"><code>E-002</code></div></li><li><p>AI can reduce correctness on tasks outside its capability frontier; consulting evidence is indirect for support. — E-003</p><div class="tribunal-evidence-refs"><code>E-003</code></div></li></ul><details class="tribunal-more"><summary>View 1 more</summary><ul><li><p>Experienced, high-skill support staff need separate quality monitoring rather than assumed uniform benefits. — E-004</p><div class="tribunal-evidence-refs"><code>E-004</code></div></li></ul></details></article><article class="tribunal-card uncertain"><header><span aria-hidden="true">?</span><h3>Cannot yet claim</h3><span class="tribunal-count">1</span></header><ul><li><p>Local net cost, privacy incident rates, subgroup service quality and sustained performance are not established by this selected evidence set [no direct evidence].</p></li></ul></article><article class="tribunal-card contradicted"><header><span aria-hidden="true">×</span><h3>Contradicted claims</h3><span class="tribunal-count">0</span></header><ul><li>No data.</li></ul></article><article class="tribunal-card missing"><header><span aria-hidden="true">…</span><h3>Missing evidence</h3><span class="tribunal-count">1</span></header><ul><li><p>Local net cost, privacy incident rates, subgroup service quality and sustained performance are not established by this selected evidence set [no direct evidence].</p></li></ul></article></div></div></div></section><section class="brief-block brief-action" id="brief-action-en"><header class="brief-block-header"><h2>Evidence to action</h2><p>Applicability, guardrails, stop conditions and evaluation form one executable path.</p></header><div class="brief-block-body"><div class="evidence-to-action" data-visual="evidence-to-action"><article class="action-node action-evidence"><span>Evidence</span><p class="action-node-text">AI assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome. — E-001</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-applicability"><span>Applicability</span><div class="action-node-text"><p>Start with low-risk, knowledge-base-covered queues; agents inspect and edit every proposed reply.; Stratify by tenure and baseline skill;…</p><details class="detail-expander"><summary>Expand full explanation</summary><div class="detail-body"><p>Start with low-risk, knowledge-base-covered queues; agents inspect and edit every proposed reply.; Stratify by tenure and baseline skill; retain expert discretion and measure review workload.; Minimize and redact customer data, restrict access and retention, and review supplier data-use terms before a pilot.; Escalate unfamiliar, ambiguous, sensitive or high-stakes requests to qualified humans; no autonomous commitments.</p></div></details></div></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-decision"><span>Decision</span><p class="action-node-text">Pilot</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-guardrails"><span>Guardrails</span><p class="action-node-text">Start with low-risk, knowledge-base-covered queues; agents inspect and edit every proposed reply.</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-stop"><span>Stop conditions</span><div class="action-node-text"><p>Pause immediately after a verified privacy leak or a serious unsafe or unauthorized customer commitment; investigate before resuming.;…</p><details class="detail-expander"><summary>Expand full explanation</summary><div class="detail-body"><p>Pause immediately after a verified privacy leak or a serious unsafe or unauthorized customer commitment; investigate before resuming.; Pause expansion if blinded quality review crosses a pre-agreed noninferiority margin, including within experience strata.</p></div></details></div></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-evaluation"><span>Evaluation</span><div class="action-node-text"><p>Expand only if quality is noninferior, verified resolutions per paid hour improve, net cost is acceptable and no serious unresolved safety incident re…</p><details class="detail-expander"><summary>Expand full explanation</summary><div class="detail-body"><p>Expand only if quality is noninferior, verified resolutions per paid hour improve, net cost is acceptable and no serious unresolved safety incident remains; thresholds require local agreement.</p></div></details></div></article></div></div></section><section class="brief-block brief-lieflat" id="brief-lieflat-en"><header class="brief-block-header"><h2>Lieflat Editorial Gallery</h2><p>Charts selected and composed by AI from the Lieflat catalog; every number traces back to result.json.</p></header><div class="brief-block-body"><div class="lieflat-gallery-container"><figure class="lieflat-card" data-lieflat data-visual="lieflat-bubble_almanac" data-chart-id="lieflat-bubble-almanac.svg"><h3 class="lieflat-title">Year × dimension evidence almanac</h3><p class="lieflat-sub">Bubble area ∝ study count · solid core = significant results</p><div class="lieflat-figure"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 540 300" width="100%" height="100%" role="img" aria-label="Year × dimension evidence almanac" style="background:#FAFAFA;">
+</div></div></section><section class="brief-block brief-outcomes" id="brief-outcomes-en"><header class="brief-block-header"><h2>Process output ≠ policy effect</h2><p>Only informative outcome separation; positive, negative and null effects use effect_direction.</p></header><div class="brief-block-body"><div class="outcome-separation" data-visual="outcome-separation"><div class="visual-heading"><h3>Outcome Separation · Process output ≠ policy effect</h3><p>Outcomes are adjudicated separately so pilot throughput or efficiency gains are not silently treated as real policy effects.</p><p class="semantic-note">Effect direction comes from evidence.effect_direction; supporting a claim does not imply a positive outcome.</p></div><div class="outcome-groups"><article class="outcome-group outcome-learning"><h3>Effect / target outcome</h3><ul><li><strong>Policy effectiveness</strong><span class="outcome-states"><span class="dir pos">Positive effect 2</span></span></li></ul></article><article class="outcome-group outcome-risk"><h3>Risk / implementation risk</h3><ul><li><strong>Implementation risk</strong><span class="outcome-states"><span class="dir neg">Negative effect 2</span></span></li></ul></article></div></div></div></section><section class="brief-block brief-tribunal" id="brief-tribunal-en"><header class="brief-block-header"><h2>Evidence tribunal</h2><p>Supported, uncertain, contradicted and missing evidence stay separated instead of flattened into long prose.</p></header><div class="brief-block-body"><div class="evidence-tribunal" data-visual="evidence-tribunal-grid"><p class="tribunal-summary"><strong>Decision: </strong>Pilot · <strong>Confidence: </strong>Moderate</p><div class="tribunal-grid"><article class="tribunal-card supported"><header><span aria-hidden="true">✓</span><h3>Can claim</h3><span class="tribunal-count">4</span></header><ul><li><p>AI assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome. — E-001</p><div class="tribunal-evidence-refs"><code>E-001</code></div></li><li><p>ChatGPT reduced time on short professional writing tasks; this is indirect evidence for customer support. — E-002</p><div class="tribunal-evidence-refs"><code>E-002</code></div></li><li><p>AI can reduce correctness on tasks outside its capability frontier; consulting evidence is indirect for support. — E-003</p><div class="tribunal-evidence-refs"><code>E-003</code></div></li></ul><details class="tribunal-more"><summary>View 1 more</summary><ul><li><p>Experienced, high-skill support staff need separate quality monitoring rather than assumed uniform benefits. — E-004</p><div class="tribunal-evidence-refs"><code>E-004</code></div></li></ul></details></article><article class="tribunal-card uncertain"><header><span aria-hidden="true">?</span><h3>Cannot yet claim</h3><span class="tribunal-count">1</span></header><ul><li><p>Local net cost, privacy incident rates, subgroup service quality and sustained performance are not established by this selected evidence set [no direct evidence].</p></li></ul></article><article class="tribunal-card contradicted"><header><span aria-hidden="true">×</span><h3>Contradicted claims</h3><span class="tribunal-count">0</span></header><ul><li>No data.</li></ul></article><article class="tribunal-card missing"><header><span aria-hidden="true">…</span><h3>Missing evidence</h3><span class="tribunal-count">1</span></header><ul><li><p>Local net cost, privacy incident rates, subgroup service quality and sustained performance are not established by this selected evidence set [no direct evidence].</p></li></ul></article></div></div></div></section><section class="brief-block brief-action" id="brief-action-en"><header class="brief-block-header"><h2>Evidence to action</h2><p>Applicability, guardrails, stop conditions and evaluation form one executable path.</p></header><div class="brief-block-body"><div class="evidence-to-action" data-visual="evidence-to-action"><article class="action-node action-evidence"><span>Evidence</span><p class="action-node-text">AI assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome. — E-001</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-applicability"><span>Applicability</span><div class="action-node-text"><p>Start with low-risk, knowledge-base-covered queues; agents inspect and edit every proposed reply.; Stratify by tenure and baseline skill;…</p><details class="detail-expander"><summary>Expand full explanation</summary><div class="detail-body"><p>Start with low-risk, knowledge-base-covered queues; agents inspect and edit every proposed reply.; Stratify by tenure and baseline skill; retain expert discretion and measure review workload.; Minimize and redact customer data, restrict access and retention, and review supplier data-use terms before a pilot.; Escalate unfamiliar, ambiguous, sensitive or high-stakes requests to qualified humans; no autonomous commitments.</p></div></details></div></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-decision"><span>Decision</span><p class="action-node-text">Pilot</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-guardrails"><span>Guardrails</span><p class="action-node-text">Start with low-risk, knowledge-base-covered queues; agents inspect and edit every proposed reply.</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-stop"><span>Stop conditions</span><div class="action-node-text"><p>Pause immediately after a verified privacy leak or a serious unsafe or unauthorized customer commitment; investigate before resuming.;…</p><details class="detail-expander"><summary>Expand full explanation</summary><div class="detail-body"><p>Pause immediately after a verified privacy leak or a serious unsafe or unauthorized customer commitment; investigate before resuming.; Pause expansion if blinded quality review crosses a pre-agreed noninferiority margin, including within experience strata.</p></div></details></div></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-evaluation"><span>Evaluation</span><div class="action-node-text"><p>Expand only if quality is noninferior, verified resolutions per paid hour improve, net cost is acceptable and no serious unresolved safety incident re…</p><details class="detail-expander"><summary>Expand full explanation</summary><div class="detail-body"><p>Expand only if quality is noninferior, verified resolutions per paid hour improve, net cost is acceptable and no serious unresolved safety incident remains; thresholds require local agreement.</p></div></details></div></article></div></div></section><section class="brief-block brief-lieflat" id="brief-lieflat-en"><header class="brief-block-header"><h2>Lieflat Editorial Gallery</h2><p>Charts selected and composed by AI from the Lieflat catalog; every number traces back to result.json.</p></header><div class="brief-block-body"><div class="lieflat-gallery-container"><figure class="lieflat-card" data-lieflat data-visual="lieflat-bubble_almanac" data-chart-id="lieflat-bubble-almanac.svg"><h3 class="lieflat-title">Year × dimension evidence almanac</h3><p class="lieflat-sub">Bubble area ∝ study count · solid core = significant results</p><div class="lieflat-figure"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 540 300" width="100%" height="100%" role="img" aria-label="Year × dimension evidence almanac" style="background:#FAFAFA;">
 <line x1="44" y1="70.0" x2="520" y2="70.0" stroke="#E2E8F0" stroke-width="0.5" class="lf-fade" style="--motion-delay:0ms"/>
 <line x1="44" y1="77.0" x2="520" y2="77.0" stroke="#E2E8F0" stroke-width="0.5" class="lf-fade" style="--motion-delay:14ms"/>
 <line x1="44" y1="84.0" x2="520" y2="84.0" stroke="#E2E8F0" stroke-width="0.5" class="lf-fade" style="--motion-delay:28ms"/>
@@ -2217,27 +2217,27 @@ select:focus-visible,
 <line x1="44" y1="245.0" x2="520" y2="245.0" stroke="#E2E8F0" stroke-width="0.5" class="lf-fade" style="--motion-delay:350ms"/>
 <line x1="44" y1="252.0" x2="520" y2="252.0" stroke="#E2E8F0" stroke-width="0.5" class="lf-fade" style="--motion-delay:364ms"/>
 <line x1="44" y1="259.0" x2="520" y2="259.0" stroke="#E2E8F0" stroke-width="0.5" class="lf-fade" style="--motion-delay:378ms"/>
-<text x="150" y="76" fill="#0F172A" font-size="9" font-weight="600" text-anchor="middle" class="lf-fade" style="--motion-delay:40ms">policy_effecti</text>
-<text x="484.8" y="76" fill="#0F172A" font-size="9" font-weight="600" text-anchor="middle" class="lf-fade" style="--motion-delay:140ms">implementation</text>
+<text x="150" y="76" fill="#0F172A" font-size="9" font-weight="600" text-anchor="middle" class="lf-fade" style="--motion-delay:40ms">Policy effecti</text>
+<text x="484.8" y="76" fill="#0F172A" font-size="9" font-weight="600" text-anchor="middle" class="lf-fade" style="--motion-delay:140ms">Implementation</text>
 <text x="96" y="96.0" fill="#475569" font-size="9" font-weight="800" text-anchor="end" class="lf-fade" style="--motion-delay:80ms">2023</text>
 <text x="96" y="140.0" fill="#475569" font-size="9" font-weight="800" text-anchor="end" class="lf-fade" style="--motion-delay:180ms">2025</text>
 <text x="96" y="184.0" fill="#475569" font-size="9" font-weight="800" text-anchor="end" class="lf-fade" style="--motion-delay:280ms">2026</text>
-<circle cx="150.0" cy="92.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:100ms"><title>policy_effectiveness (2023) — N = 1 studies, significant = 0</title></circle>
-<circle cx="150.0" cy="136.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:112ms"><title>policy_effectiveness (2025) — N = 1 studies, significant = 0</title></circle>
-<circle cx="520.0" cy="136.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:124ms"><title>implementation_risk (2025) — N = 1 studies, significant = 0</title></circle>
-<circle cx="520.0" cy="180.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:136ms"><title>implementation_risk (2026) — N = 1 studies, significant = 0</title></circle>
+<circle cx="150.0" cy="92.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:100ms"><title>Policy effectiveness (2023) — N = 1 studies, significant = 0</title></circle>
+<circle cx="150.0" cy="136.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:112ms"><title>Policy effectiveness (2025) — N = 1 studies, significant = 0</title></circle>
+<circle cx="520.0" cy="136.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:124ms"><title>Implementation risk (2025) — N = 1 studies, significant = 0</title></circle>
+<circle cx="520.0" cy="180.0" r="3.6" fill="#0F172A" fill-opacity="0.22" stroke="#0F172A" stroke-width="1.2" class="lf-pop" style="--motion-delay:136ms"><title>Implementation risk (2026) — N = 1 studies, significant = 0</title></circle>
 </svg></div><figcaption class="lieflat-caption">Drawn only when years and outcome dimensions exist.</figcaption><p class="lieflat-src">L9 Bubble Almanac · Evidence.year X Dimension</p></figure><figure class="lieflat-card" data-lieflat data-visual="lieflat-matrix_heat" data-chart-id="lieflat-matrix-heat.svg"><h3 class="lieflat-title">Year × outcome evidence density</h3><p class="lieflat-sub">Each cell counts evidence items for that year and outcome</p><div class="lieflat-figure"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 540 200" width="100%" height="100%" role="img" aria-label="Year × outcome evidence density" style="background:#FAFAFA;">
 <text x="211.7" y="66" text-anchor="middle" font-size="9" font-weight="800" fill="#475569" class="lf-fade" style="--motion-delay:40ms">2023</text>
 <text x="335.0" y="66" text-anchor="middle" font-size="9" font-weight="800" fill="#475569" class="lf-fade" style="--motion-delay:140ms">2025</text>
 <text x="458.3" y="66" text-anchor="middle" font-size="9" font-weight="800" fill="#475569" class="lf-fade" style="--motion-delay:240ms">2026</text>
-<text x="138" y="97.0" text-anchor="end" font-size="8" font-weight="600" fill="#475569" class="lf-fade" style="--motion-delay:60ms">policy effecti</text>
+<text x="138" y="97.0" text-anchor="end" font-size="8" font-weight="600" fill="#475569" class="lf-fade" style="--motion-delay:60ms">Policy effecti</text>
 <rect x="152.0" y="80.0" width="119.3" height="28" rx="4" fill="#0F172A" fill-opacity="0.90" class="lf-pop" style="--motion-delay:0ms"/>
 <text x="211.7" y="97.0" text-anchor="middle" font-size="8.0" font-weight="800" fill="#0F172A" class="lf-pop" style="--motion-delay:60ms">1</text>
 <rect x="275.3" y="80.0" width="119.3" height="28" rx="4" fill="#0F172A" fill-opacity="0.90" class="lf-pop" style="--motion-delay:12ms"/>
 <text x="335.0" y="97.0" text-anchor="middle" font-size="8.0" font-weight="800" fill="#0F172A" class="lf-pop" style="--motion-delay:72ms">1</text>
 <rect x="398.7" y="80.0" width="119.3" height="28" rx="4" fill="#E2E8F0" fill-opacity="0.5" class="lf-fade" style="--motion-delay:24ms"/>
 <text x="458.3" y="97.0" text-anchor="middle" font-size="8.0" font-weight="800" fill="#0F172A" class="lf-pop" style="--motion-delay:84ms">0</text>
-<text x="138" y="129.0" text-anchor="end" font-size="8" font-weight="600" fill="#475569" class="lf-fade" style="--motion-delay:160ms">implementation</text>
+<text x="138" y="129.0" text-anchor="end" font-size="8" font-weight="600" fill="#475569" class="lf-fade" style="--motion-delay:160ms">Implementation</text>
 <rect x="152.0" y="112.0" width="119.3" height="28" rx="4" fill="#E2E8F0" fill-opacity="0.5" class="lf-fade" style="--motion-delay:12ms"/>
 <text x="211.7" y="129.0" text-anchor="middle" font-size="8.0" font-weight="800" fill="#0F172A" class="lf-pop" style="--motion-delay:72ms">0</text>
 <rect x="275.3" y="112.0" width="119.3" height="28" rx="4" fill="#0F172A" fill-opacity="0.90" class="lf-pop" style="--motion-delay:24ms"/>
@@ -2394,7 +2394,7 @@ select:focus-visible,
 </svg></div><figcaption class="lieflat-caption">Shows which checklist items concentrate the weaknesses.</figcaption><p class="lieflat-src">L15 Ballot Tally · Methodology.flag Rates</p></figure></div></div></section><section class="brief-block brief-sources" id="brief-sources-en"><header class="brief-block-header"><h2>Key sources</h2><p>Only the key sources in the brief; full traceability expands in the full report.</p></header><div class="brief-block-body"><div class="brief-source-grid"><article class="brief-source"><code>S-001</code><h3><a href="https://academic.oup.com/qje/article/140/2/889/7990658">Generative AI at Work</a></h3><p>Tier 1 DOI-verified paper · 2025</p></article><article class="brief-source"><code>S-002</code><h3><a href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf">Experimental evidence on the productivity effects of generative artificial intelligence</a></h3><p>Tier 1 DOI-verified paper · 2023</p></article><article class="brief-source"><code>S-003</code><h3><a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838">Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality</a></h3><p>Tier 1 DOI-verified paper · 2026</p></article></div></div></section></div></div>
 <div class="report-page report-page-full" data-report-page="full" hidden>
 <div class="full-report-intro"><h2>Full Report</h2><p>Conclusions first: every traceable piece of evidence and method note lives here, with visuals only at points where they add meaning. Every number traces back to result.json.</p></div>
-<div class="full-report-layout"><aside class="full-report-toc" aria-label="Contents"><div class="toc-head"><strong>Contents</strong><button type="button" class="toc-collapse" aria-expanded="true" data-label-collapse="Collapse contents" data-label-expand="Expand contents">Collapse contents</button></div><nav><a href="#full-01-decision-en" data-toc-target="full-01-decision-en" data-chapter-key="decision">01 Decision, Adjudication &amp; Research Boundary</a><a href="#full-02-evidence-en" data-toc-target="full-02-evidence-en" data-chapter-key="evidence">02 Key Evidence &amp; Outcome Separation</a><a href="#full-03-quality-en" data-toc-target="full-03-quality-en" data-chapter-key="quality">03 Evidence Quality, Counterevidence &amp; Method Audit</a><a href="#full-04-action-en" data-toc-target="full-04-action-en" data-chapter-key="action">04 Applicability &amp; Teaching Action</a><a href="#full-05-evaluation-en" data-toc-target="full-05-evaluation-en" data-chapter-key="evaluation">05 Pilot, Evaluation &amp; Stop Conditions</a><a href="#full-06-sources-en" data-toc-target="full-06-sources-en" data-chapter-key="sources">06 Sources, Traceability &amp; Appendix</a></nav></aside><main class="full-report-content"><section id="full-01-decision-en" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>01 Decision, Adjudication &amp; Research Boundary</h2><p class="full-chapter-lead">State the final adjudication and research boundary before explaining why.</p></header><div class="full-chapter-body">
+<div class="full-report-layout"><aside class="full-report-toc" aria-label="Contents"><div class="toc-head"><strong>Contents</strong><button type="button" class="toc-collapse" aria-expanded="true" data-label-collapse="Collapse contents" data-label-expand="Expand contents">Collapse contents</button></div><nav><a href="#full-01-decision-en" data-toc-target="full-01-decision-en" data-chapter-key="decision">01 Decision, Adjudication &amp; Research Boundary</a><a href="#full-02-evidence-en" data-toc-target="full-02-evidence-en" data-chapter-key="evidence">02 Key Evidence &amp; Outcome Separation</a><a href="#full-03-quality-en" data-toc-target="full-03-quality-en" data-chapter-key="quality">03 Evidence Quality, Counterevidence &amp; Method Audit</a><a href="#full-04-action-en" data-toc-target="full-04-action-en" data-chapter-key="action">04 Applicability &amp; Policy Action</a><a href="#full-05-evaluation-en" data-toc-target="full-05-evaluation-en" data-chapter-key="evaluation">05 Pilot, Evaluation &amp; Stop Conditions</a><a href="#full-06-sources-en" data-toc-target="full-06-sources-en" data-chapter-key="sources">06 Sources, Traceability &amp; Appendix</a></nav></aside><main class="full-report-content"><section id="full-01-decision-en" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>01 Decision, Adjudication &amp; Research Boundary</h2><p class="full-chapter-lead">State the final adjudication and research boundary before explaining why.</p></header><div class="full-chapter-body">
 <div class="decision-hero pilot" data-visual="decision-hero">
   <div class="hero-decision">
     <span class="eyebrow">Recommended decision</span>
@@ -2409,19 +2409,19 @@ select:focus-visible,
     <article class="hero-insight next"><span>Next action</span><p class="hero-insight-text">Run a supervised pilot on approved knowledge bases with human review on every reply, and track escalation and correction rates.</p></article>
   </div>
   <p class="hero-provenance"><span>Evidence / sources</span> · 4 / 3</p>
-</div><div class="scope-grid"><article class="scope-card"><h3>Research question</h3><p class="scope-text">Should an enterprise customer-support team introduce a generative AI assistant?</p></article><article class="scope-card"><h3>AI intervention</h3><p class="scope-text">Policy: Human Supervised Customer Support Assistant; Policy type: Institutional reform; Mechanism: Approved knowledge base + agent review</p></article><article class="scope-card"><h3>Comparison</h3><p class="scope-text">Existing support workflow without generative AI suggestions.</p></article><article class="scope-card"><h3>Outcome constructs</h3><p class="scope-text">Primary outcomes: Policy Effectiveness, Implementation Risk; Secondary outcomes: Cost Effectiveness, equity, feasibility</p></article><article class="scope-card"><h3>Research scope</h3><p class="scope-text">Time range: 2023–2026; Evidence Types: Quasi-experimental, Randomised controlled trial</p></article><article class="scope-card"><h3>Decision success condition</h3><p class="scope-text">Improve verified resolutions per paid staff hour while preserving service quality, privacy and staff autonomy.</p></article></div></div></section><section id="full-02-evidence-en" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>02 Key Evidence &amp; Outcome Separation</h2><p class="full-chapter-lead">Place task performance, actual learning, retention and risk on one evidence map without conflating them.</p></header><div class="full-chapter-body"><div class="retrieval-grid"><article><h3>Inclusion criteria</h3><ul></ul></article><article><h3>Exclusion criteria</h3><ul></ul></article></div><div class="retrieval-coverage"><h3>Source coverage</h3><p><code>S-001</code> <code>S-002</code> <code>S-003</code></p><p class="retrieval-note">This report shows only retrieval metadata present in result; it does not fabricate PRISMA/funnel counts when none exist.</p></div><div class="outcome-separation" data-visual="outcome-separation"><div class="visual-heading"><h3>Outcome Separation · Task performance ≠ learning</h3><p>Outcomes are adjudicated separately so faster training performance is not silently treated as evidence of learning.</p><p class="semantic-note">Effect direction comes from evidence.effect_direction; supporting a claim does not imply a positive outcome.</p></div><div class="outcome-groups"><article class="outcome-group outcome-other"><h3>Other outcomes</h3><ul><li><strong>Policy Effectiveness</strong><span class="outcome-states"><span class="dir pos">Positive effect 2</span></span></li><li><strong>Implementation Risk</strong><span class="outcome-states"><span class="dir neg">Negative effect 2</span></span></li></ul></article></div></div><div class='table-wrap outcome-table'><table class='data-table'><thead><tr><th>Outcome</th><th>Positive effect</th><th>Negative effect</th><th>Null effect</th><th>Evidence</th></tr></thead><tbody><tr><td><strong>Policy Effectiveness</strong><span class='raw-tag' title='raw id'>policy_effectiveness</span></td><td class='num'>2</td><td class='num'>0</td><td class='num'>0</td><td><code>E-001</code> <code>E-002</code> </td></tr><tr><td><strong>Implementation Risk</strong><span class='raw-tag' title='raw id'>implementation_risk</span></td><td class='num'>0</td><td class='num'>2</td><td class='num'>0</td><td><code>E-003</code> <code>E-004</code> </td></tr></tbody></table></div><div class='matrix-controls'><div class='matrix-tools'><input id='matrix-search-full-en' type='search' placeholder='Search evidence…' aria-label='Filter / search evidence'><select id='matrix-direction-full-en' aria-label='Filter by effect direction'><option value=''>All effects</option><option value='positive'>Positive effect</option><option value='negative'>Negative effect</option><option value='null'>Null effect</option></select><select id='matrix-outcome-full-en' aria-label='Filter by outcome type'><option value=''>All outcomes</option><option value='implementation_risk'>Implementation Risk</option><option value='policy_effectiveness'>Policy Effectiveness</option></select></div></div><div class='table-wrap matrix-wrap'><table id='evidence-matrix-full-en' class='data-table evidence-matrix'><thead><tr><th>ID</th><th>Outcome</th><th>Effect</th><th>Quality</th><th>Claim</th><th>Source</th></tr></thead><tbody><tr data-effect="positive" data-direction="positive" data-outcome="policy_effectiveness" data-search="e-001 st-001 sample-support-all generative ai at work ai assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome. customer-support agents   positive s-001"><td><code>E-001</code></td><td><strong>Policy Effectiveness</strong></td><td><span class="dir pos">Positive effect</span><span class="relation-note">Claim relation: Support</span></td><td><div class="quality-cell"><span class="num">9</span><span class="quality-meter" aria-hidden="true"><i style="width:90%"></i></span></div></td><td class="claim-cell"><p>AI assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome.</p><details class="matrix-row-detail"><summary>View full evidence</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>Study ID</dt><dd>ST-001</dd></div><div class="evidence-detail-row"><dt>Sample ID</dt><dd>SAMPLE-support-all</dd></div><div class="evidence-detail-row"><dt>Study title</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>Source title</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>Year</dt><dd>2025</dd></div><div class="evidence-detail-row"><dt>Study design</dt><dd>Quasi-experimental</dd></div><div class="evidence-detail-row"><dt>Population</dt><dd>Customer-support agents</dd></div><div class="evidence-detail-row"><dt>Sample size</dt><dd>5172</dd></div><div class="evidence-detail-row"><dt>Outcome measure</dt><dd>Chat handling time; separate throughput summary: about 15% more issues resolved per hour.</dd></div><div class="evidence-detail-row"><dt>Effect direction</dt><dd>Positive effect</dd></div><div class="evidence-detail-row"><dt>Relation to claim</dt><dd>Support</dd></div><div class="evidence-detail-row"><dt>Limitations</dt><dd>One firm, one tool and a staggered nonrandom rollout; causal interpretation depends on identification assumptions.</dd></div><div class="evidence-detail-row"><dt>Quality dimensions</dt><dd>D1 Study design = 1 · D2 Sample quality = 2 · D3 Measurement validity = 2 · D4 Temporal strength = 2 · D5 Directness = 2</dd></div><div class="evidence-detail-row"><dt>Quality score</dt><dd>9.0</dd></div><div class="evidence-detail-row"><dt>Evidence status</dt><dd>Supported</dd></div><div class="evidence-detail-row"><dt>Full claim</dt><dd>AI assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome.</dd></div><div class="evidence-detail-row"><dt>Source location</dt><dd>https://academic.oup.com/qje/article/140/2/889/7990658</dd></div><div class="evidence-detail-row"><dt>Verifiable link</dt><dd><a href="https://academic.oup.com/qje/article/140/2/889/7990658">https://academic.oup.com/qje/article/140/2/889/7990658</a></dd></div></dl></details></td><td><a class="source-link" href="https://academic.oup.com/qje/article/140/2/889/7990658" title="Generative AI at Work"><code>S-001</code></a></td></tr><tr data-effect="positive" data-direction="positive" data-outcome="policy_effectiveness" data-search="e-002 st-002 sample-writing experimental evidence on the productivity effects of generative artificial intelligence chatgpt reduced time on short professional writing tasks; this is indirect evidence for customer support. college-educated working professionals   positive s-002"><td><code>E-002</code></td><td><strong>Policy Effectiveness</strong></td><td><span class="dir pos">Positive effect</span><span class="relation-note">Claim relation: Support</span></td><td><div class="quality-cell"><span class="num">7</span><span class="quality-meter" aria-hidden="true"><i style="width:70%"></i></span></div></td><td class="claim-cell"><p>ChatGPT reduced time on short professional writing tasks; this is indirect evidence for customer support.</p><details class="matrix-row-detail"><summary>View full evidence</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>Study ID</dt><dd>ST-002</dd></div><div class="evidence-detail-row"><dt>Sample ID</dt><dd>SAMPLE-writing</dd></div><div class="evidence-detail-row"><dt>Study title</dt><dd>Experimental evidence on the productivity effects of generative artificial intelligence</dd></div><div class="evidence-detail-row"><dt>Source title</dt><dd>Experimental evidence on the productivity effects of generative artificial intelligence</dd></div><div class="evidence-detail-row"><dt>Year</dt><dd>2023</dd></div><div class="evidence-detail-row"><dt>Study design</dt><dd>Randomized controlled trial</dd></div><div class="evidence-detail-row"><dt>Population</dt><dd>College-educated working professionals</dd></div><div class="evidence-detail-row"><dt>Sample size</dt><dd>453</dd></div><div class="evidence-detail-row"><dt>Outcome measure</dt><dd>Self-reported task time: approximately 40% lower; assessed writing quality approximately 18% higher is a separate measure.</dd></div><div class="evidence-detail-row"><dt>Effect direction</dt><dd>Positive effect</dd></div><div class="evidence-detail-row"><dt>Relation to claim</dt><dd>Support</dd></div><div class="evidence-detail-row"><dt>Limitations</dt><dd>Indirect evidence, not directly generalizable: brief incentivized writing tasks did not demand precise factual accuracy or customer-specific context.</dd></div><div class="evidence-detail-row"><dt>Quality dimensions</dt><dd>D1 Study design = 2 · D2 Sample quality = 2 · D3 Measurement validity = 1 · D4 Temporal strength = 1 · D5 Directness = 1</dd></div><div class="evidence-detail-row"><dt>Quality score</dt><dd>7.0</dd></div><div class="evidence-detail-row"><dt>Evidence status</dt><dd>Supported</dd></div><div class="evidence-detail-row"><dt>Full claim</dt><dd>ChatGPT reduced time on short professional writing tasks; this is indirect evidence for customer support.</dd></div><div class="evidence-detail-row"><dt>Source location</dt><dd>https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</dd></div><div class="evidence-detail-row"><dt>Verifiable link</dt><dd><a href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf">https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</a></dd></div></dl></details></td><td><a class="source-link" href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf" title="Experimental evidence on the productivity effects of generative artificial intelligence"><code>S-002</code></a></td></tr><tr data-effect="negative" data-direction="negative" data-outcome="implementation_risk" data-search="e-003 st-003 sample-consulting-outside navigating the jagged technological frontier: field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality ai can reduce correctness on tasks outside its capability frontier; consulting evidence is indirect for support. bcg consultants in the outside-frontier experiment   negative s-003"><td><code>E-003</code></td><td><strong>Implementation Risk</strong></td><td><span class="dir neg">Negative effect</span><span class="relation-note">Claim relation: Support</span></td><td><div class="quality-cell"><span class="num">8</span><span class="quality-meter" aria-hidden="true"><i style="width:80%"></i></span></div></td><td class="claim-cell"><p>AI can reduce correctness on tasks outside its capability frontier; consulting evidence is indirect for support.</p><details class="matrix-row-detail"><summary>View full evidence</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>Study ID</dt><dd>ST-003</dd></div><div class="evidence-detail-row"><dt>Sample ID</dt><dd>SAMPLE-consulting-outside</dd></div><div class="evidence-detail-row"><dt>Study title</dt><dd>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality</dd></div><div class="evidence-detail-row"><dt>Source title</dt><dd>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality</dd></div><div class="evidence-detail-row"><dt>Year</dt><dd>2026</dd></div><div class="evidence-detail-row"><dt>Study design</dt><dd>Randomized controlled trial</dd></div><div class="evidence-detail-row"><dt>Population</dt><dd>BCG consultants in the outside-frontier experiment</dd></div><div class="evidence-detail-row"><dt>Sample size</dt><dd>373</dd></div><div class="evidence-detail-row"><dt>Outcome measure</dt><dd>Correct business recommendation: about 19 percentage points lower across AI arms; Table 7 has 373 participants, not 758.</dd></div><div class="evidence-detail-row"><dt>Effect direction</dt><dd>Negative effect</dd></div><div class="evidence-detail-row"><dt>Relation to claim</dt><dd>Support</dd></div><div class="evidence-detail-row"><dt>Limitations</dt><dd>Indirect evidence, not directly generalizable: consultants solving an experimental business case, not live customer tickets.</dd></div><div class="evidence-detail-row"><dt>Quality dimensions</dt><dd>D1 Study design = 2 · D2 Sample quality = 2 · D3 Measurement validity = 2 · D4 Temporal strength = 1 · D5 Directness = 1</dd></div><div class="evidence-detail-row"><dt>Quality score</dt><dd>8.0</dd></div><div class="evidence-detail-row"><dt>Evidence status</dt><dd>Supported</dd></div><div class="evidence-detail-row"><dt>Full claim</dt><dd>AI can reduce correctness on tasks outside its capability frontier; consulting evidence is indirect for support.</dd></div><div class="evidence-detail-row"><dt>Source location</dt><dd>https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</dd></div><div class="evidence-detail-row"><dt>Verifiable link</dt><dd><a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838">https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</a></dd></div></dl></details></td><td><a class="source-link" href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838" title="Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality"><code>S-003</code></a></td></tr><tr data-effect="negative" data-direction="negative" data-outcome="implementation_risk" data-search="e-004 st-001 sample-support-all generative ai at work experienced, high-skill support staff need separate quality monitoring rather than assumed uniform benefits. experienced and high-skill customer-support agents   negative s-001"><td><code>E-004</code></td><td><strong>Implementation Risk</strong></td><td><span class="dir neg">Negative effect</span><span class="relation-note">Claim relation: Support</span></td><td><div class="quality-cell"><span class="num">8</span><span class="quality-meter" aria-hidden="true"><i style="width:80%"></i></span></div></td><td class="claim-cell"><p>Experienced, high-skill support staff need separate quality monitoring rather than assumed uniform benefits.</p><details class="matrix-row-detail"><summary>View full evidence</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>Study ID</dt><dd>ST-001</dd></div><div class="evidence-detail-row"><dt>Sample ID</dt><dd>SAMPLE-support-all</dd></div><div class="evidence-detail-row"><dt>Study title</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>Source title</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>Year</dt><dd>2025</dd></div><div class="evidence-detail-row"><dt>Study design</dt><dd>Quasi-experimental</dd></div><div class="evidence-detail-row"><dt>Population</dt><dd>Experienced and high-skill customer-support agents</dd></div><div class="evidence-detail-row"><dt>Outcome measure</dt><dd>Small quality declines among the most experienced and highest-skilled support staff.</dd></div><div class="evidence-detail-row"><dt>Effect direction</dt><dd>Negative effect</dd></div><div class="evidence-detail-row"><dt>Relation to claim</dt><dd>Support</dd></div><div class="evidence-detail-row"><dt>Limitations</dt><dd>Same study as E-001; subgroup sample size was not extracted. This is not an independent replication.</dd></div><div class="evidence-detail-row"><dt>Quality dimensions</dt><dd>D1 Study design = 1 · D2 Sample quality = 1 · D3 Measurement validity = 2 · D4 Temporal strength = 2 · D5 Directness = 2</dd></div><div class="evidence-detail-row"><dt>Quality score</dt><dd>8.0</dd></div><div class="evidence-detail-row"><dt>Evidence status</dt><dd>Supported</dd></div><div class="evidence-detail-row"><dt>Full claim</dt><dd>Experienced, high-skill support staff need separate quality monitoring rather than assumed uniform benefits.</dd></div><div class="evidence-detail-row"><dt>Source location</dt><dd>https://academic.oup.com/qje/article/140/2/889/7990658</dd></div><div class="evidence-detail-row"><dt>Verifiable link</dt><dd><a href="https://academic.oup.com/qje/article/140/2/889/7990658">https://academic.oup.com/qje/article/140/2/889/7990658</a></dd></div></dl></details></td><td><a class="source-link" href="https://academic.oup.com/qje/article/140/2/889/7990658" title="Generative AI at Work"><code>S-001</code></a></td></tr></tbody></table></div></div></section><section id="full-03-quality-en" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>03 Evidence Quality, Counterevidence &amp; Method Audit</h2><p class="full-chapter-lead">Examine why evidence is credible, where it conflicts, and which conclusions require downgrading.</p></header><div class="full-chapter-body"><section class="method-review-group"><header class="method-review-title"><h3>Audit target: overall</h3><span class="method-verdict">Concern</span></header><div class="method-audit-grid"><article class="method-audit-item method-met"><div class="method-audit-head"><strong>Evidence Level</strong><span class="method-status">Met</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>Causal Identification</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>External Validity</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-missing"><div class="method-audit-head"><strong>Cost Evidence</strong><span class="method-status">missing</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>Stakeholder Representation</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>Implementation Evidence</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>Equity Analysis</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>Effect Size Reported</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>Uncertainty Quantified</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-met"><div class="method-audit-head"><strong>Comparator Clarity</strong><span class="method-status">Met</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>Publication Bias Risk</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-met"><div class="method-audit-head"><strong>Generalizability Claims</strong><span class="method-status">Met</span></div><p></p></article></div><div class="method-guard-wrap"><strong>Task vs learning guard: </strong><p class="method-guard">Teaching outcomes are not applicable; productivity is not a claim about student learning.</p></div></section><p>No data.</p><div class="evidence-tribunal" data-visual="evidence-tribunal-grid"><p class="tribunal-summary"><strong>Decision: </strong>Pilot · <strong>Confidence: </strong>Moderate</p><div class="tribunal-grid"><article class="tribunal-card supported"><header><span aria-hidden="true">✓</span><h3>Can claim</h3><span class="tribunal-count">4</span></header><ul><li><p>AI assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome. — E-001</p><div class="tribunal-evidence-refs"><code>E-001</code></div></li><li><p>ChatGPT reduced time on short professional writing tasks; this is indirect evidence for customer support. — E-002</p><div class="tribunal-evidence-refs"><code>E-002</code></div></li><li><p>AI can reduce correctness on tasks outside its capability frontier; consulting evidence is indirect for support. — E-003</p><div class="tribunal-evidence-refs"><code>E-003</code></div></li><li><p>Experienced, high-skill support staff need separate quality monitoring rather than assumed uniform benefits. — E-004</p><div class="tribunal-evidence-refs"><code>E-004</code></div></li></ul></article><article class="tribunal-card uncertain"><header><span aria-hidden="true">?</span><h3>Cannot yet claim</h3><span class="tribunal-count">1</span></header><ul><li><p>Local net cost, privacy incident rates, subgroup service quality and sustained performance are not established by this selected evidence set [no direct evidence].</p></li></ul></article><article class="tribunal-card contradicted"><header><span aria-hidden="true">×</span><h3>Contradicted claims</h3><span class="tribunal-count">0</span></header><ul><li>No data.</li></ul></article><article class="tribunal-card missing"><header><span aria-hidden="true">…</span><h3>Missing evidence</h3><span class="tribunal-count">1</span></header><ul><li><p>Local net cost, privacy incident rates, subgroup service quality and sustained performance are not established by this selected evidence set [no direct evidence].</p></li></ul></article></div><details class="supporting-visual"><summary>EvidenceFlow Protocol</summary><svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="EvidenceFlow Protocol"><title>EvidenceFlow Protocol</title><desc>Research flow from framing, retrieval, fetch/verify, extraction, challenge, method audit and adjudication to applicability and intervention evaluation.</desc><defs><marker id="arr-en-full-workflow" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">EvidenceFlow Protocol</text><rect x="24" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="56.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="56.0" y="126.0">Frame</tspan></text><line x1="88" y1="126" x2="100" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="100" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="132.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="132.0" y="126.0">Retrieve</tspan></text><line x1="164" y1="126" x2="176" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="176" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="208.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="208.0" y="126.0">Fetch</tspan><tspan x="208.0" y="140.0">Verify</tspan></text><line x1="240" y1="126" x2="252" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="252" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="284.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="284.0" y="126.0">Extract</tspan></text><line x1="316" y1="126" x2="328" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="328" y="100" width="64" height="52" rx="8" fill="#5E8A6A"/><text x="360.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="360.0" y="126.0">Challenge</tspan></text><line x1="392" y1="126" x2="404" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="404" y="100" width="64" height="52" rx="8" fill="#5E8A6A"/><text x="436.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="436.0" y="126.0">Audit</tspan><tspan x="436.0" y="140.0">Method</tspan></text><line x1="468" y1="126" x2="480" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="480" y="100" width="64" height="52" rx="8" fill="#C99A4A"/><text x="512.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="512.0" y="126.0">Adjudicate</tspan></text><line x1="544" y1="126" x2="556" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="556" y="100" width="64" height="52" rx="8" fill="#C99A4A"/><text x="588.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="588.0" y="126.0">Applicability</tspan></text><line x1="620" y1="126" x2="632" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="632" y="100" width="64" height="52" rx="8" fill="#C99A4A"/><text x="664.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="664.0" y="126.0">Intervene</tspan><tspan x="664.0" y="140.0">Evaluate</tspan></text></svg></details><details class="supporting-visual"><summary>Tribunal infographic</summary><svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Evidence Tribunal infographic"><title>Evidence Tribunal infographic</title><desc>Evidence IDs for claims that can and cannot be claimed, plus the recommended action badge; full claim text is in the tribunal cards below.</desc><defs><marker id="arr-en-full-tribunal" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">Evidence Tribunal</text><text x="24" y="180" font-size="13" font-weight="700" fill="#3A3833">Source of conflict</text><text x="24" y="202" font-size="11" fill="#8A867E">See tribunal cards below</text><rect x="24" y="216" width="180" height="28" rx="14" fill="#C99A4A"/><text x="114" y="235" text-anchor="middle" font-size="13" font-weight="700" fill="#fff">Pilot</text><text x="250" y="80" font-size="13" font-weight="700" fill="#5E8A6A">Can claim (4)</text><circle cx="254" cy="100" r="3" fill="#5E8A6A"/><text x="266" y="105" font-size="11" fill="#3A3833">E-001</text><circle cx="254" cy="122" r="3" fill="#5E8A6A"/><text x="266" y="127" font-size="11" fill="#3A3833">E-002</text><circle cx="254" cy="144" r="3" fill="#5E8A6A"/><text x="266" y="149" font-size="11" fill="#3A3833">E-003</text><circle cx="254" cy="166" r="3" fill="#5E8A6A"/><text x="266" y="171" font-size="11" fill="#3A3833">E-004</text><text x="490" y="80" font-size="13" font-weight="700" fill="#A85B53">Cannot claim (1)</text><circle cx="494" cy="100" r="3" fill="#A85B53"/><text x="506" y="105" font-size="11" fill="#3A3833">(none)</text></svg></details></div><div class="trace-chain" data-visual="trace-chain"><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-001</strong><p class="trace-claim-text">AI assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome.</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-001</code><span>Policy Effectiveness</span><span class="dir pos">Positive effect</span><span class="trace-relation">claim relation: Support</span><span>Q 9.0</span><span class="trace-arrow">→</span><a href="https://academic.oup.com/qje/article/140/2/889/7990658"><code>S-001</code></a></div></div></article><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-002</strong><p class="trace-claim-text">ChatGPT reduced time on short professional writing tasks; this is indirect evidence for customer support.</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-002</code><span>Policy Effectiveness</span><span class="dir pos">Positive effect</span><span class="trace-relation">claim relation: Support</span><span>Q 7.0</span><span class="trace-arrow">→</span><a href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf"><code>S-002</code></a></div></div></article><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-003</strong><p class="trace-claim-text">AI can reduce correctness on tasks outside its capability frontier; consulting evidence is indirect for support.</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-003</code><span>Implementation Risk</span><span class="dir neg">Negative effect</span><span class="trace-relation">claim relation: Support</span><span>Q 8.0</span><span class="trace-arrow">→</span><a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838"><code>S-003</code></a></div></div></article><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-004</strong><p class="trace-claim-text">Experienced, high-skill support staff need separate quality monitoring rather than assumed uniform benefits.</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-004</code><span>Implementation Risk</span><span class="dir neg">Negative effect</span><span class="trace-relation">claim relation: Support</span><span>Q 8.0</span><span class="trace-arrow">→</span><a href="https://academic.oup.com/qje/article/140/2/889/7990658"><code>S-001</code></a></div></div></article></div><div id="chart-trace-en" class="chart-mount" aria-label="Claim-Evidence Trace"></div><p class="chart-interpretation"><strong>What this means: </strong>Every important claim must resolve to Evidence IDs and original sources.</p></div></section><section id="full-04-action-en" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>04 Applicability &amp; Teaching Action</h2><p class="full-chapter-lead">Connect applicability, guardrails and teaching actions to specific evidence.</p></header><div class="full-chapter-body"><p><strong>Target population: </strong>Enterprise customer-support staff, stratified by tenure and baseline skill.</p>
+</div><div class="scope-grid"><article class="scope-card"><h3>Research question</h3><p class="scope-text">Should an enterprise customer-support team introduce a generative AI assistant?</p></article><article class="scope-card"><h3>Intervention plan</h3><p class="scope-text">Policy: Human-supervised customer-support assistant; Policy type: Institutional reform; Mechanism: Approved knowledge base + agent review</p></article><article class="scope-card"><h3>Comparison</h3><p class="scope-text">Existing support workflow without generative AI suggestions.</p></article><article class="scope-card"><h3>Outcome constructs</h3><p class="scope-text">Primary outcomes: Policy effectiveness, Implementation risk; Secondary outcomes: Cost-effectiveness, Equity, Feasibility</p></article><article class="scope-card"><h3>Research scope</h3><p class="scope-text">Time range: 2023–2026; Evidence Types: Quasi-experimental, Randomised controlled trial</p></article><article class="scope-card"><h3>Decision success condition</h3><p class="scope-text">Improve verified resolutions per paid staff hour while preserving service quality, privacy and staff autonomy.</p></article></div></div></section><section id="full-02-evidence-en" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>02 Key Evidence &amp; Outcome Separation</h2><p class="full-chapter-lead">Place process outputs, real effects, durability and risk on one evidence map without conflating them.</p></header><div class="full-chapter-body"><div class="retrieval-grid"><article><h3>Inclusion criteria</h3><ul></ul></article><article><h3>Exclusion criteria</h3><ul></ul></article></div><div class="retrieval-coverage"><h3>Source coverage</h3><p><code>S-001</code> <code>S-002</code> <code>S-003</code></p><p class="retrieval-note">This report shows only retrieval metadata present in result; it does not fabricate PRISMA/funnel counts when none exist.</p></div><div class="outcome-separation" data-visual="outcome-separation"><div class="visual-heading"><h3>Outcome Separation · Process output ≠ policy effect</h3><p>Outcomes are adjudicated separately so pilot throughput or efficiency gains are not silently treated as real policy effects.</p><p class="semantic-note">Effect direction comes from evidence.effect_direction; supporting a claim does not imply a positive outcome.</p></div><div class="outcome-groups"><article class="outcome-group outcome-learning"><h3>Effect / target outcome</h3><ul><li><strong>Policy effectiveness</strong><span class="outcome-states"><span class="dir pos">Positive effect 2</span></span></li></ul></article><article class="outcome-group outcome-risk"><h3>Risk / implementation risk</h3><ul><li><strong>Implementation risk</strong><span class="outcome-states"><span class="dir neg">Negative effect 2</span></span></li></ul></article></div></div><div class='table-wrap outcome-table'><table class='data-table'><thead><tr><th>Outcome</th><th>Positive effect</th><th>Negative effect</th><th>Null effect</th><th>Evidence</th></tr></thead><tbody><tr><td><strong>Policy effectiveness</strong><span class='raw-tag' title='raw id'>policy_effectiveness</span></td><td class='num'>2</td><td class='num'>0</td><td class='num'>0</td><td><code>E-001</code> <code>E-002</code> </td></tr><tr><td><strong>Implementation risk</strong><span class='raw-tag' title='raw id'>implementation_risk</span></td><td class='num'>0</td><td class='num'>2</td><td class='num'>0</td><td><code>E-003</code> <code>E-004</code> </td></tr></tbody></table></div><div class='matrix-controls'><div class='matrix-tools'><input id='matrix-search-full-en' type='search' placeholder='Search evidence…' aria-label='Filter / search evidence'><select id='matrix-direction-full-en' aria-label='Filter by effect direction'><option value=''>All effects</option><option value='positive'>Positive effect</option><option value='negative'>Negative effect</option><option value='null'>Null effect</option></select><select id='matrix-outcome-full-en' aria-label='Filter by outcome type'><option value=''>All outcomes</option><option value='implementation_risk'>Implementation risk</option><option value='policy_effectiveness'>Policy effectiveness</option></select></div></div><div class='table-wrap matrix-wrap'><table id='evidence-matrix-full-en' class='data-table evidence-matrix'><thead><tr><th>ID</th><th>Outcome</th><th>Effect</th><th>Quality</th><th>Claim</th><th>Source</th></tr></thead><tbody><tr data-effect="positive" data-direction="positive" data-outcome="policy_effectiveness" data-search="e-001 st-001 sample-support-all generative ai at work ai assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome. customer-support agents   positive s-001"><td><code>E-001</code></td><td><strong>Policy effectiveness</strong></td><td><span class="dir pos">Positive effect</span><span class="relation-note">Claim relation: Support</span></td><td><div class="quality-cell"><span class="num">9</span><span class="quality-meter" aria-hidden="true"><i style="width:90%"></i></span></div></td><td class="claim-cell"><p>AI assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome.</p><details class="matrix-row-detail"><summary>View full evidence</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>Study ID</dt><dd>ST-001</dd></div><div class="evidence-detail-row"><dt>Sample ID</dt><dd>SAMPLE-support-all</dd></div><div class="evidence-detail-row"><dt>Study title</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>Source title</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>Year</dt><dd>2025</dd></div><div class="evidence-detail-row"><dt>Study design</dt><dd>Quasi-experimental</dd></div><div class="evidence-detail-row"><dt>Population</dt><dd>Customer-support agents</dd></div><div class="evidence-detail-row"><dt>Sample size</dt><dd>5172</dd></div><div class="evidence-detail-row"><dt>Outcome measure</dt><dd>Chat handling time; separate throughput summary: about 15% more issues resolved per hour.</dd></div><div class="evidence-detail-row"><dt>Effect direction</dt><dd>Positive effect</dd></div><div class="evidence-detail-row"><dt>Relation to claim</dt><dd>Support</dd></div><div class="evidence-detail-row"><dt>Limitations</dt><dd>One firm, one tool and a staggered nonrandom rollout; causal interpretation depends on identification assumptions.</dd></div><div class="evidence-detail-row"><dt>Quality dimensions</dt><dd>D1 Study design = 1 · D2 Sample quality = 2 · D3 Measurement validity = 2 · D4 Temporal strength = 2 · D5 Directness = 2</dd></div><div class="evidence-detail-row"><dt>Quality score</dt><dd>9.0</dd></div><div class="evidence-detail-row"><dt>Evidence status</dt><dd>Supported</dd></div><div class="evidence-detail-row"><dt>Full claim</dt><dd>AI assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome.</dd></div><div class="evidence-detail-row"><dt>Source location</dt><dd>https://academic.oup.com/qje/article/140/2/889/7990658</dd></div><div class="evidence-detail-row"><dt>Verifiable URL</dt><dd><a href="https://academic.oup.com/qje/article/140/2/889/7990658">https://academic.oup.com/qje/article/140/2/889/7990658</a></dd></div></dl></details></td><td><a class="source-link" href="https://academic.oup.com/qje/article/140/2/889/7990658" title="Generative AI at Work"><code>S-001</code></a></td></tr><tr data-effect="positive" data-direction="positive" data-outcome="policy_effectiveness" data-search="e-002 st-002 sample-writing experimental evidence on the productivity effects of generative artificial intelligence chatgpt reduced time on short professional writing tasks; this is indirect evidence for customer support. college-educated working professionals   positive s-002"><td><code>E-002</code></td><td><strong>Policy effectiveness</strong></td><td><span class="dir pos">Positive effect</span><span class="relation-note">Claim relation: Support</span></td><td><div class="quality-cell"><span class="num">7</span><span class="quality-meter" aria-hidden="true"><i style="width:70%"></i></span></div></td><td class="claim-cell"><p>ChatGPT reduced time on short professional writing tasks; this is indirect evidence for customer support.</p><details class="matrix-row-detail"><summary>View full evidence</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>Study ID</dt><dd>ST-002</dd></div><div class="evidence-detail-row"><dt>Sample ID</dt><dd>SAMPLE-writing</dd></div><div class="evidence-detail-row"><dt>Study title</dt><dd>Experimental evidence on the productivity effects of generative artificial intelligence</dd></div><div class="evidence-detail-row"><dt>Source title</dt><dd>Experimental evidence on the productivity effects of generative artificial intelligence</dd></div><div class="evidence-detail-row"><dt>Year</dt><dd>2023</dd></div><div class="evidence-detail-row"><dt>Study design</dt><dd>Randomized controlled trial</dd></div><div class="evidence-detail-row"><dt>Population</dt><dd>College-educated working professionals</dd></div><div class="evidence-detail-row"><dt>Sample size</dt><dd>453</dd></div><div class="evidence-detail-row"><dt>Outcome measure</dt><dd>Self-reported task time: approximately 40% lower; assessed writing quality approximately 18% higher is a separate measure.</dd></div><div class="evidence-detail-row"><dt>Effect direction</dt><dd>Positive effect</dd></div><div class="evidence-detail-row"><dt>Relation to claim</dt><dd>Support</dd></div><div class="evidence-detail-row"><dt>Limitations</dt><dd>Indirect evidence, not directly generalizable: brief incentivized writing tasks did not demand precise factual accuracy or customer-specific context.</dd></div><div class="evidence-detail-row"><dt>Quality dimensions</dt><dd>D1 Study design = 2 · D2 Sample quality = 2 · D3 Measurement validity = 1 · D4 Temporal strength = 1 · D5 Directness = 1</dd></div><div class="evidence-detail-row"><dt>Quality score</dt><dd>7.0</dd></div><div class="evidence-detail-row"><dt>Evidence status</dt><dd>Supported</dd></div><div class="evidence-detail-row"><dt>Full claim</dt><dd>ChatGPT reduced time on short professional writing tasks; this is indirect evidence for customer support.</dd></div><div class="evidence-detail-row"><dt>Source location</dt><dd>https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</dd></div><div class="evidence-detail-row"><dt>Verifiable URL</dt><dd><a href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf">https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</a></dd></div></dl></details></td><td><a class="source-link" href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf" title="Experimental evidence on the productivity effects of generative artificial intelligence"><code>S-002</code></a></td></tr><tr data-effect="negative" data-direction="negative" data-outcome="implementation_risk" data-search="e-003 st-003 sample-consulting-outside navigating the jagged technological frontier: field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality ai can reduce correctness on tasks outside its capability frontier; consulting evidence is indirect for support. bcg consultants in the outside-frontier experiment   negative s-003"><td><code>E-003</code></td><td><strong>Implementation risk</strong></td><td><span class="dir neg">Negative effect</span><span class="relation-note">Claim relation: Support</span></td><td><div class="quality-cell"><span class="num">8</span><span class="quality-meter" aria-hidden="true"><i style="width:80%"></i></span></div></td><td class="claim-cell"><p>AI can reduce correctness on tasks outside its capability frontier; consulting evidence is indirect for support.</p><details class="matrix-row-detail"><summary>View full evidence</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>Study ID</dt><dd>ST-003</dd></div><div class="evidence-detail-row"><dt>Sample ID</dt><dd>SAMPLE-consulting-outside</dd></div><div class="evidence-detail-row"><dt>Study title</dt><dd>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality</dd></div><div class="evidence-detail-row"><dt>Source title</dt><dd>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality</dd></div><div class="evidence-detail-row"><dt>Year</dt><dd>2026</dd></div><div class="evidence-detail-row"><dt>Study design</dt><dd>Randomized controlled trial</dd></div><div class="evidence-detail-row"><dt>Population</dt><dd>BCG consultants in the outside-frontier experiment</dd></div><div class="evidence-detail-row"><dt>Sample size</dt><dd>373</dd></div><div class="evidence-detail-row"><dt>Outcome measure</dt><dd>Correct business recommendation: about 19 percentage points lower across AI arms; Table 7 has 373 participants, not 758.</dd></div><div class="evidence-detail-row"><dt>Effect direction</dt><dd>Negative effect</dd></div><div class="evidence-detail-row"><dt>Relation to claim</dt><dd>Support</dd></div><div class="evidence-detail-row"><dt>Limitations</dt><dd>Indirect evidence, not directly generalizable: consultants solving an experimental business case, not live customer tickets.</dd></div><div class="evidence-detail-row"><dt>Quality dimensions</dt><dd>D1 Study design = 2 · D2 Sample quality = 2 · D3 Measurement validity = 2 · D4 Temporal strength = 1 · D5 Directness = 1</dd></div><div class="evidence-detail-row"><dt>Quality score</dt><dd>8.0</dd></div><div class="evidence-detail-row"><dt>Evidence status</dt><dd>Supported</dd></div><div class="evidence-detail-row"><dt>Full claim</dt><dd>AI can reduce correctness on tasks outside its capability frontier; consulting evidence is indirect for support.</dd></div><div class="evidence-detail-row"><dt>Source location</dt><dd>https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</dd></div><div class="evidence-detail-row"><dt>Verifiable URL</dt><dd><a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838">https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</a></dd></div></dl></details></td><td><a class="source-link" href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838" title="Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality"><code>S-003</code></a></td></tr><tr data-effect="negative" data-direction="negative" data-outcome="implementation_risk" data-search="e-004 st-001 sample-support-all generative ai at work experienced, high-skill support staff need separate quality monitoring rather than assumed uniform benefits. experienced and high-skill customer-support agents   negative s-001"><td><code>E-004</code></td><td><strong>Implementation risk</strong></td><td><span class="dir neg">Negative effect</span><span class="relation-note">Claim relation: Support</span></td><td><div class="quality-cell"><span class="num">8</span><span class="quality-meter" aria-hidden="true"><i style="width:80%"></i></span></div></td><td class="claim-cell"><p>Experienced, high-skill support staff need separate quality monitoring rather than assumed uniform benefits.</p><details class="matrix-row-detail"><summary>View full evidence</summary><dl class="evidence-detail-grid"><div class="evidence-detail-row"><dt>Study ID</dt><dd>ST-001</dd></div><div class="evidence-detail-row"><dt>Sample ID</dt><dd>SAMPLE-support-all</dd></div><div class="evidence-detail-row"><dt>Study title</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>Source title</dt><dd>Generative AI at Work</dd></div><div class="evidence-detail-row"><dt>Year</dt><dd>2025</dd></div><div class="evidence-detail-row"><dt>Study design</dt><dd>Quasi-experimental</dd></div><div class="evidence-detail-row"><dt>Population</dt><dd>Experienced and high-skill customer-support agents</dd></div><div class="evidence-detail-row"><dt>Outcome measure</dt><dd>Small quality declines among the most experienced and highest-skilled support staff.</dd></div><div class="evidence-detail-row"><dt>Effect direction</dt><dd>Negative effect</dd></div><div class="evidence-detail-row"><dt>Relation to claim</dt><dd>Support</dd></div><div class="evidence-detail-row"><dt>Limitations</dt><dd>Same study as E-001; subgroup sample size was not extracted. This is not an independent replication.</dd></div><div class="evidence-detail-row"><dt>Quality dimensions</dt><dd>D1 Study design = 1 · D2 Sample quality = 1 · D3 Measurement validity = 2 · D4 Temporal strength = 2 · D5 Directness = 2</dd></div><div class="evidence-detail-row"><dt>Quality score</dt><dd>8.0</dd></div><div class="evidence-detail-row"><dt>Evidence status</dt><dd>Supported</dd></div><div class="evidence-detail-row"><dt>Full claim</dt><dd>Experienced, high-skill support staff need separate quality monitoring rather than assumed uniform benefits.</dd></div><div class="evidence-detail-row"><dt>Source location</dt><dd>https://academic.oup.com/qje/article/140/2/889/7990658</dd></div><div class="evidence-detail-row"><dt>Verifiable URL</dt><dd><a href="https://academic.oup.com/qje/article/140/2/889/7990658">https://academic.oup.com/qje/article/140/2/889/7990658</a></dd></div></dl></details></td><td><a class="source-link" href="https://academic.oup.com/qje/article/140/2/889/7990658" title="Generative AI at Work"><code>S-001</code></a></td></tr></tbody></table></div></div></section><section id="full-03-quality-en" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>03 Evidence Quality, Counterevidence &amp; Method Audit</h2><p class="full-chapter-lead">Examine why evidence is credible, where it conflicts, and which conclusions require downgrading.</p></header><div class="full-chapter-body"><section class="method-review-group"><header class="method-review-title"><h3>Audit target: overall</h3><span class="method-verdict">Concern</span></header><div class="method-audit-grid"><article class="method-audit-item method-met"><div class="method-audit-head"><strong>Evidence level</strong><span class="method-status">Met</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>Causal identification</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>External validity</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-missing"><div class="method-audit-head"><strong>Cost evidence</strong><span class="method-status">missing</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>Stakeholder representation</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>Implementation evidence</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>Equity analysis</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>Effect size reported</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>Uncertainty quantified</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-met"><div class="method-audit-head"><strong>Comparator clarity</strong><span class="method-status">Met</span></div><p></p></article><article class="method-audit-item method-partial"><div class="method-audit-head"><strong>Publication-bias risk</strong><span class="method-status">Partial</span></div><p></p></article><article class="method-audit-item method-met"><div class="method-audit-head"><strong>Generalizability claims</strong><span class="method-status">Met</span></div><p></p></article></div><div class="method-guard-wrap"><strong>Output vs effect guard: </strong><p class="method-guard">Teaching outcomes are not applicable; productivity is not a claim about student learning.</p></div></section><p>No data.</p><div class="evidence-tribunal" data-visual="evidence-tribunal-grid"><p class="tribunal-summary"><strong>Decision: </strong>Pilot · <strong>Confidence: </strong>Moderate</p><div class="tribunal-grid"><article class="tribunal-card supported"><header><span aria-hidden="true">✓</span><h3>Can claim</h3><span class="tribunal-count">4</span></header><ul><li><p>AI assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome. — E-001</p><div class="tribunal-evidence-refs"><code>E-001</code></div></li><li><p>ChatGPT reduced time on short professional writing tasks; this is indirect evidence for customer support. — E-002</p><div class="tribunal-evidence-refs"><code>E-002</code></div></li><li><p>AI can reduce correctness on tasks outside its capability frontier; consulting evidence is indirect for support. — E-003</p><div class="tribunal-evidence-refs"><code>E-003</code></div></li><li><p>Experienced, high-skill support staff need separate quality monitoring rather than assumed uniform benefits. — E-004</p><div class="tribunal-evidence-refs"><code>E-004</code></div></li></ul></article><article class="tribunal-card uncertain"><header><span aria-hidden="true">?</span><h3>Cannot yet claim</h3><span class="tribunal-count">1</span></header><ul><li><p>Local net cost, privacy incident rates, subgroup service quality and sustained performance are not established by this selected evidence set [no direct evidence].</p></li></ul></article><article class="tribunal-card contradicted"><header><span aria-hidden="true">×</span><h3>Contradicted claims</h3><span class="tribunal-count">0</span></header><ul><li>No data.</li></ul></article><article class="tribunal-card missing"><header><span aria-hidden="true">…</span><h3>Missing evidence</h3><span class="tribunal-count">1</span></header><ul><li><p>Local net cost, privacy incident rates, subgroup service quality and sustained performance are not established by this selected evidence set [no direct evidence].</p></li></ul></article></div><details class="supporting-visual"><summary>EvidenceFlow Protocol</summary><svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="EvidenceFlow Protocol"><title>EvidenceFlow Protocol</title><desc>Research flow from framing, retrieval, fetch/verify, extraction, challenge, method audit and adjudication to applicability and intervention evaluation.</desc><defs><marker id="arr-en-full-workflow" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">EvidenceFlow Protocol</text><rect x="24" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="56.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="56.0" y="126.0">Frame</tspan></text><line x1="88" y1="126" x2="100" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="100" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="132.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="132.0" y="126.0">Retrieve</tspan></text><line x1="164" y1="126" x2="176" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="176" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="208.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="208.0" y="126.0">Fetch</tspan><tspan x="208.0" y="140.0">Verify</tspan></text><line x1="240" y1="126" x2="252" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="252" y="100" width="64" height="52" rx="8" fill="#B8694A"/><text x="284.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="284.0" y="126.0">Extract</tspan></text><line x1="316" y1="126" x2="328" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="328" y="100" width="64" height="52" rx="8" fill="#5E8A6A"/><text x="360.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="360.0" y="126.0">Challenge</tspan></text><line x1="392" y1="126" x2="404" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="404" y="100" width="64" height="52" rx="8" fill="#5E8A6A"/><text x="436.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="436.0" y="126.0">Audit</tspan><tspan x="436.0" y="140.0">Method</tspan></text><line x1="468" y1="126" x2="480" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="480" y="100" width="64" height="52" rx="8" fill="#C99A4A"/><text x="512.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="512.0" y="126.0">Adjudicate</tspan></text><line x1="544" y1="126" x2="556" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="556" y="100" width="64" height="52" rx="8" fill="#C99A4A"/><text x="588.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="588.0" y="126.0">Applicability</tspan></text><line x1="620" y1="126" x2="632" y2="126" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-workflow)"/><rect x="632" y="100" width="64" height="52" rx="8" fill="#C99A4A"/><text x="664.0" y="126.0" text-anchor="middle" fill="#FFFFFF" font-size="10" font-weight="600"><tspan x="664.0" y="126.0">Intervene</tspan><tspan x="664.0" y="140.0">Evaluate</tspan></text></svg></details><details class="supporting-visual"><summary>Tribunal infographic</summary><svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Evidence Tribunal infographic"><title>Evidence Tribunal infographic</title><desc>Evidence IDs for claims that can and cannot be claimed, plus the recommended action badge; full claim text is in the tribunal cards below.</desc><defs><marker id="arr-en-full-tribunal" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">Evidence Tribunal</text><text x="24" y="180" font-size="13" font-weight="700" fill="#3A3833">Source of conflict</text><text x="24" y="202" font-size="11" fill="#8A867E">See tribunal cards below</text><rect x="24" y="216" width="180" height="28" rx="14" fill="#C99A4A"/><text x="114" y="235" text-anchor="middle" font-size="13" font-weight="700" fill="#fff">Pilot</text><text x="250" y="80" font-size="13" font-weight="700" fill="#5E8A6A">Can claim (4)</text><circle cx="254" cy="100" r="3" fill="#5E8A6A"/><text x="266" y="105" font-size="11" fill="#3A3833">E-001</text><circle cx="254" cy="122" r="3" fill="#5E8A6A"/><text x="266" y="127" font-size="11" fill="#3A3833">E-002</text><circle cx="254" cy="144" r="3" fill="#5E8A6A"/><text x="266" y="149" font-size="11" fill="#3A3833">E-003</text><circle cx="254" cy="166" r="3" fill="#5E8A6A"/><text x="266" y="171" font-size="11" fill="#3A3833">E-004</text><text x="490" y="80" font-size="13" font-weight="700" fill="#A85B53">Cannot claim (1)</text><circle cx="494" cy="100" r="3" fill="#A85B53"/><text x="506" y="105" font-size="11" fill="#3A3833">(none)</text></svg></details></div><div class="trace-chain" data-visual="trace-chain"><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-001</strong><p class="trace-claim-text">AI assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome.</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-001</code><span>Policy effectiveness</span><span class="dir pos">Positive effect</span><span class="trace-relation">claim relation: Support</span><span>Q 9.0</span><span class="trace-arrow">→</span><a href="https://academic.oup.com/qje/article/140/2/889/7990658"><code>S-001</code></a></div></div></article><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-002</strong><p class="trace-claim-text">ChatGPT reduced time on short professional writing tasks; this is indirect evidence for customer support.</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-002</code><span>Policy effectiveness</span><span class="dir pos">Positive effect</span><span class="trace-relation">claim relation: Support</span><span>Q 7.0</span><span class="trace-arrow">→</span><a href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf"><code>S-002</code></a></div></div></article><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-003</strong><p class="trace-claim-text">AI can reduce correctness on tasks outside its capability frontier; consulting evidence is indirect for support.</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-003</code><span>Implementation risk</span><span class="dir neg">Negative effect</span><span class="trace-relation">claim relation: Support</span><span>Q 8.0</span><span class="trace-arrow">→</span><a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838"><code>S-003</code></a></div></div></article><article class="trace-chain-card"><div class="trace-claim-node"><strong>C-004</strong><p class="trace-claim-text">Experienced, high-skill support staff need separate quality monitoring rather than assumed uniform benefits.</p></div><div class="trace-arrow">↓</div><div class="trace-evidence-list"><div class="trace-evidence-node"><code>E-004</code><span>Implementation risk</span><span class="dir neg">Negative effect</span><span class="trace-relation">claim relation: Support</span><span>Q 8.0</span><span class="trace-arrow">→</span><a href="https://academic.oup.com/qje/article/140/2/889/7990658"><code>S-001</code></a></div></div></article></div><div id="chart-trace-en" class="chart-mount" aria-label="Claim-Evidence Trace"></div><p class="chart-interpretation"><strong>What this means: </strong>Every important claim must resolve to Evidence IDs and original sources.</p></div></section><section id="full-04-action-en" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>04 Applicability &amp; Policy Action</h2><p class="full-chapter-lead">Connect applicability, guardrails and policy actions to specific evidence.</p></header><div class="full-chapter-body"><p><strong>Target population: </strong>Enterprise customer-support staff, stratified by tenure and baseline skill.</p>
 <p><strong>Target context: </strong>Human-supervised support using an approved knowledge base.</p>
 <p><strong>Conditions: </strong></p><ul><li>Start with low-risk, knowledge-base-covered queues; agents inspect and edit every proposed reply.</li><li>Stratify by tenure and baseline skill; retain expert discretion and measure review workload.</li><li>Minimize and redact customer data, restrict access and retention, and review supplier data-use terms before a pilot.</li><li>Escalate unfamiliar, ambiguous, sensitive or high-stakes requests to qualified humans; no autonomous commitments.</li></ul><div class="boundary-block"><h3>Claims beyond the evidence boundary</h3><ul><li>Claiming universal gains or that autonomous deployment is safe exceeds the boundary: no included study measures privacy incidents, local net cost or subgroup service quality.</li></ul></div><div class="evidence-to-action" data-visual="evidence-to-action"><article class="action-node action-evidence"><span>Evidence</span><p class="action-node-text">AI assistance can shorten customer chat handling in a bounded support setting; throughput is a separate outcome. — E-001</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-applicability"><span>Applicability</span><div class="action-node-text"><p>Start with low-risk, knowledge-base-covered queues; agents inspect and edit every proposed reply.; Stratify by tenure and baseline skill;…</p><details class="detail-expander"><summary>Expand full explanation</summary><div class="detail-body"><p>Start with low-risk, knowledge-base-covered queues; agents inspect and edit every proposed reply.; Stratify by tenure and baseline skill; retain expert discretion and measure review workload.; Minimize and redact customer data, restrict access and retention, and review supplier data-use terms before a pilot.; Escalate unfamiliar, ambiguous, sensitive or high-stakes requests to qualified humans; no autonomous commitments.</p></div></details></div></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-decision"><span>Decision</span><p class="action-node-text">Pilot</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-guardrails"><span>Guardrails</span><p class="action-node-text">Start with low-risk, knowledge-base-covered queues; agents inspect and edit every proposed reply.</p></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-stop"><span>Stop conditions</span><div class="action-node-text"><p>Pause immediately after a verified privacy leak or a serious unsafe or unauthorized customer commitment; investigate before resuming.;…</p><details class="detail-expander"><summary>Expand full explanation</summary><div class="detail-body"><p>Pause immediately after a verified privacy leak or a serious unsafe or unauthorized customer commitment; investigate before resuming.; Pause expansion if blinded quality review crosses a pre-agreed noninferiority margin, including within experience strata.</p></div></details></div></article><span class="flow-arrow" aria-hidden="true">→</span><article class="action-node action-evaluation"><span>Evaluation</span><div class="action-node-text"><p>Expand only if quality is noninferior, verified resolutions per paid hour improve, net cost is acceptable and no serious unresolved safety incident re…</p><details class="detail-expander"><summary>Expand full explanation</summary><div class="detail-body"><p>Expand only if quality is noninferior, verified resolutions per paid hour improve, net cost is acceptable and no serious unresolved safety incident remains; thresholds require local agreement.</p></div></details></div></article></div><p><strong>Target population: </strong>Enterprise customer-support staff, stratified by tenure and baseline skill. · <strong>Pilot duration: </strong>Proposed: two baseline weeks and six pilot weeks; not executed.</p>
-<p><strong>AI usage policy: </strong>Start with low-risk, knowledge-base-covered queues; agents inspect and edit every proposed reply.</p>
+<p><strong>Intervention usage policy: </strong>Start with low-risk, knowledge-base-covered queues; agents inspect and edit every proposed reply.</p>
 <h3>Stop conditions</h3><ul><li>Pause immediately after a verified privacy leak or a serious unsafe or unauthorized customer commitment; investigate before resuming.</li><li>Pause expansion if blinded quality review crosses a pre-agreed noninferiority margin, including within experience strata.</li></ul>
-<h3>Intervention timeline infographic</h3>
-<svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Teaching intervention timeline"><title>Teaching intervention timeline</title><desc>Short phase names and activity counts; full AI usage rules are in the phase blocks.</desc><defs><marker id="arr-en-full-intervention" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">Intervention Timeline</text><rect x="24" y="100" width="160" height="96" rx="8" fill="#B8694A"/><text x="104.0" y="148.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="104.0" y="148.0">Phase 1</tspan></text></svg></div></section><section id="full-05-evaluation-en" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>05 Pilot, Evaluation &amp; Stop Conditions</h2><p class="full-chapter-lead">Validate the pilot with independent learning outcomes and pre-specified stop conditions.</p></header><div class="full-chapter-body"><p><strong>Research question: </strong>Should an enterprise customer-support team introduce a generative AI assistant?</p>
+<h3>Intervention plan timeline infographic</h3>
+<svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Intervention plan timeline"><title>Intervention plan timeline</title><desc>Short phase names and activity counts; full intervention usage rules are in the phase blocks.</desc><defs><marker id="arr-en-full-intervention" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">Intervention plan timeline</text><rect x="24" y="100" width="160" height="96" rx="8" fill="#B8694A"/><text x="104.0" y="148.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="104.0" y="148.0">Phase 1</tspan></text></svg></div></section><section id="full-05-evaluation-en" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>05 Pilot, Evaluation &amp; Stop Conditions</h2><p class="full-chapter-lead">Validate the pilot with independent effect measures and pre-specified stop conditions.</p></header><div class="full-chapter-body"><p><strong>Research question: </strong>Should an enterprise customer-support team introduce a generative AI assistant?</p>
 <p><strong>Baseline: </strong>Record resolution rate, paid hours, repeat contacts, blinded quality and review costs before allocation.</p>
-<p><strong>Post test: </strong>Assess the same outcomes at pilot end; separately audit privacy and unsafe commitments.</p>
+<p><strong>Pilot end: </strong>Assess the same outcomes at pilot end; separately audit privacy and unsafe commitments.</p>
 <p><strong>Success threshold: </strong>Expand only if quality is noninferior, verified resolutions per paid hour improve, net cost is acceptable and no serious unresolved safety incident remains; thresholds require local agreement.</p>
 <p><strong>Analysis plan: </strong>Preregister an intention-to-treat comparison with team-clustered uncertainty and subgroup estimates. Set sample size and noninferiority margins from local baseline variance and operational priorities before enrollment.</p>
 <h3>Evaluation design infographic</h3>
-<svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Evaluation design flow"><title>Evaluation design flow</title><desc>Evaluation flow across baseline, post test, retention and transfer; full metrics and analysis plan are in the evaluation section.</desc><defs><marker id="arr-en-full-evaluation" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">Evaluation Design Flow</text><rect x="30" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="105.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="105.0" y="138.0">Baseline</tspan></text><text x="105.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">pre-test</text><line x1="180" y1="138" x2="192" y2="138" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-evaluation)"/><rect x="192" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="267.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="267.0" y="138.0">Post test</tspan></text><text x="267.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">post-test</text><line x1="342" y1="138" x2="354" y2="138" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-evaluation)"/><rect x="354" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="429.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="429.0" y="138.0">Retention</tspan></text><text x="429.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">retention test</text><line x1="504" y1="138" x2="516" y2="138" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-evaluation)"/><rect x="516" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="591.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="591.0" y="138.0">Transfer</tspan></text><text x="591.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">transfer test (no AI)</text></svg><div class="visual-suppressed"><strong>Benchmark visual suppressed</strong><p>result.json carries no benchmark.baselines, so this visual is omitted; see the standalone benchmark report.</p></div></div></section><section id="full-06-sources-en" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>06 Sources, Traceability &amp; Appendix</h2><p class="full-chapter-lead">Preserve original sources, URLs, evidence IDs and retrieval metadata for auditability.</p></header><div class="full-chapter-body"><h3>Source list</h3><div class='table-wrap'><table class='data-table source-table'><thead><tr><th>ID</th><th>Title</th><th>Year</th><th>Authority</th><th>Verifiable location</th></tr></thead><tbody><tr><td><code>S-001</code></td><td class='cell-main source-title-cell'>Generative AI at Work<details class="source-expander detail-expander"><summary>View source &amp; provenance</summary><dl class="source-detail-grid"><div class="source-detail-row"><dt>Original title</dt><dd>Generative AI at Work</dd></div><div class="source-detail-row"><dt>Year</dt><dd>2025</dd></div><div class="source-detail-row"><dt>Authority</dt><dd>Tier 1 DOI-verified paper</dd></div><div class="source-detail-row"><dt>Fetch provider</dt><dd>builtin</dd></div><div class="source-detail-row"><dt>Fetch status</dt><dd>FETCH_VALID</dd></div><div class="source-detail-row"><dt>Verifiable link</dt><dd><a href="https://academic.oup.com/qje/article/140/2/889/7990658">https://academic.oup.com/qje/article/140/2/889/7990658</a></dd></div></dl></details></td><td>2025</td><td>Tier 1 DOI-verified paper</td><td class='cell-main'><a href='https://academic.oup.com/qje/article/140/2/889/7990658'>https://academic.oup.com/qje/article/140/2/889/7990658</a></td></tr><tr><td><code>S-002</code></td><td class='cell-main source-title-cell'>Experimental evidence on the productivity effects of generative artificial intelligence<details class="source-expander detail-expander"><summary>View source &amp; provenance</summary><dl class="source-detail-grid"><div class="source-detail-row"><dt>Original title</dt><dd>Experimental evidence on the productivity effects of generative artificial intelligence</dd></div><div class="source-detail-row"><dt>Year</dt><dd>2023</dd></div><div class="source-detail-row"><dt>Authority</dt><dd>Tier 1 DOI-verified paper</dd></div><div class="source-detail-row"><dt>Fetch provider</dt><dd>builtin</dd></div><div class="source-detail-row"><dt>Fetch status</dt><dd>FETCH_VALID</dd></div><div class="source-detail-row"><dt>Verifiable link</dt><dd><a href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf">https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</a></dd></div></dl></details></td><td>2023</td><td>Tier 1 DOI-verified paper</td><td class='cell-main'><a href='https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf'>https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</a></td></tr><tr><td><code>S-003</code></td><td class='cell-main source-title-cell'>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality<details class="source-expander detail-expander"><summary>View source &amp; provenance</summary><dl class="source-detail-grid"><div class="source-detail-row"><dt>Original title</dt><dd>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality</dd></div><div class="source-detail-row"><dt>Year</dt><dd>2026</dd></div><div class="source-detail-row"><dt>Authority</dt><dd>Tier 1 DOI-verified paper</dd></div><div class="source-detail-row"><dt>Fetch provider</dt><dd>builtin</dd></div><div class="source-detail-row"><dt>Fetch status</dt><dd>FETCH_VALID</dd></div><div class="source-detail-row"><dt>Verifiable link</dt><dd><a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838">https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</a></dd></div></dl></details></td><td>2026</td><td>Tier 1 DOI-verified paper</td><td class='cell-main'><a href='https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838'>https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</a></td></tr></tbody></table></div><h3>Fetch provenance</h3><div class='table-wrap'><table class='data-table'><thead><tr><th>Source</th><th>Fetch method</th><th>Status</th><th>Fallback</th><th>Time</th></tr></thead><tbody><tr><td><code>S-001</code></td><td>builtin</td><td>FETCH_VALID</td><td></td><td></td></tr><tr><td><code>S-002</code></td><td>builtin</td><td>FETCH_VALID</td><td></td><td></td></tr><tr><td><code>S-003</code></td><td>builtin</td><td>FETCH_VALID</td><td></td><td></td></tr></tbody></table></div></div></section></main></div>
+<svg viewBox="0 0 720 260" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Evaluation design flow"><title>Evaluation design flow</title><desc>Evaluation flow across baseline, pilot end, durability and scale-out; full metrics and analysis plan are in the evaluation section.</desc><defs><marker id="arr-en-full-evaluation" markerWidth="8" markerHeight="8" refX="6" refY="3" orient="auto"><path d="M0,0 L6,3 L0,6 Z" fill="#8A867E"/></marker></defs><rect width="720" height="260" fill="#FCFAF6" rx="12"/><text x="24" y="34" font-size="16" font-weight="700" fill="#3A3833">Evaluation design flow</text><rect x="30" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="105.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="105.0" y="138.0">Baseline</tspan></text><text x="105.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">pre-test</text><line x1="180" y1="138" x2="192" y2="138" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-evaluation)"/><rect x="192" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="267.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="267.0" y="138.0">Post test</tspan></text><text x="267.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">post-test</text><line x1="342" y1="138" x2="354" y2="138" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-evaluation)"/><rect x="354" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="429.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="429.0" y="138.0">Retention</tspan></text><text x="429.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">Durability</text><line x1="504" y1="138" x2="516" y2="138" stroke="#8A867E" stroke-width="2" marker-end="url(#arr-en-full-evaluation)"/><rect x="516" y="110" width="150" height="56" rx="8" fill="#5E8A6A"/><text x="591.0" y="138.0" text-anchor="middle" fill="#FFFFFF" font-size="12" font-weight="600"><tspan x="591.0" y="138.0">Transfer</tspan></text><text x="591.0" y="148" text-anchor="middle" font-size="9" fill="#fff" opacity="0.95">Scale-out</text></svg><div class="visual-suppressed"><strong>Benchmark visual suppressed</strong><p>result.json carries no benchmark.baselines, so this visual is omitted; see the standalone benchmark report.</p></div></div></section><section id="full-06-sources-en" class="full-chapter" data-full-chapter><header class="full-chapter-header"><h2>06 Sources, Traceability &amp; Appendix</h2><p class="full-chapter-lead">Preserve original sources, URLs, evidence IDs and retrieval metadata for auditability.</p></header><div class="full-chapter-body"><h3>Source list</h3><div class='table-wrap'><table class='data-table source-table'><thead><tr><th>ID</th><th>Title</th><th>Year</th><th>Authority</th><th>Verifiable location</th></tr></thead><tbody><tr><td><code>S-001</code></td><td class='cell-main source-title-cell'>Generative AI at Work<details class="source-expander detail-expander"><summary>View source &amp; provenance</summary><dl class="source-detail-grid"><div class="source-detail-row"><dt>Original title</dt><dd>Generative AI at Work</dd></div><div class="source-detail-row"><dt>Year</dt><dd>2025</dd></div><div class="source-detail-row"><dt>Authority</dt><dd>Tier 1 DOI-verified paper</dd></div><div class="source-detail-row"><dt>Fetch provider</dt><dd>builtin</dd></div><div class="source-detail-row"><dt>Fetch status</dt><dd>FETCH_VALID</dd></div><div class="source-detail-row"><dt>Verifiable link</dt><dd><a href="https://academic.oup.com/qje/article/140/2/889/7990658">https://academic.oup.com/qje/article/140/2/889/7990658</a></dd></div></dl></details></td><td>2025</td><td>Tier 1 DOI-verified paper</td><td class='cell-main'><a href='https://academic.oup.com/qje/article/140/2/889/7990658'>https://academic.oup.com/qje/article/140/2/889/7990658</a></td></tr><tr><td><code>S-002</code></td><td class='cell-main source-title-cell'>Experimental evidence on the productivity effects of generative artificial intelligence<details class="source-expander detail-expander"><summary>View source &amp; provenance</summary><dl class="source-detail-grid"><div class="source-detail-row"><dt>Original title</dt><dd>Experimental evidence on the productivity effects of generative artificial intelligence</dd></div><div class="source-detail-row"><dt>Year</dt><dd>2023</dd></div><div class="source-detail-row"><dt>Authority</dt><dd>Tier 1 DOI-verified paper</dd></div><div class="source-detail-row"><dt>Fetch provider</dt><dd>builtin</dd></div><div class="source-detail-row"><dt>Fetch status</dt><dd>FETCH_VALID</dd></div><div class="source-detail-row"><dt>Verifiable link</dt><dd><a href="https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf">https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</a></dd></div></dl></details></td><td>2023</td><td>Tier 1 DOI-verified paper</td><td class='cell-main'><a href='https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf'>https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf</a></td></tr><tr><td><code>S-003</code></td><td class='cell-main source-title-cell'>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality<details class="source-expander detail-expander"><summary>View source &amp; provenance</summary><dl class="source-detail-grid"><div class="source-detail-row"><dt>Original title</dt><dd>Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality</dd></div><div class="source-detail-row"><dt>Year</dt><dd>2026</dd></div><div class="source-detail-row"><dt>Authority</dt><dd>Tier 1 DOI-verified paper</dd></div><div class="source-detail-row"><dt>Fetch provider</dt><dd>builtin</dd></div><div class="source-detail-row"><dt>Fetch status</dt><dd>FETCH_VALID</dd></div><div class="source-detail-row"><dt>Verifiable link</dt><dd><a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838">https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</a></dd></div></dl></details></td><td>2026</td><td>Tier 1 DOI-verified paper</td><td class='cell-main'><a href='https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838'>https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838</a></td></tr></tbody></table></div><h3>Fetch provenance</h3><div class='table-wrap'><table class='data-table'><thead><tr><th>Source</th><th>Fetch method</th><th>Status</th><th>Fallback</th><th>Time</th></tr></thead><tbody><tr><td><code>S-001</code></td><td>builtin</td><td>FETCH_VALID</td><td></td><td></td></tr><tr><td><code>S-002</code></td><td>builtin</td><td>FETCH_VALID</td><td></td><td></td></tr><tr><td><code>S-003</code></td><td>builtin</td><td>FETCH_VALID</td><td></td><td></td></tr></tbody></table></div></div></section></main></div>
 </div>
 <footer class="report-footer"><p>EduEvidence Evidence Report · Schema PASS · Claim Binding PASS · Numeric Consistency PASS · Bilingual Structure PASS · Human Language PASS · False Precision PASS · Lieflat Data Bound PASS · Axis Distortion NOT_CHECKED · Colorblind Safe NOT_CHECKED · single-file offline · source: result.json</p></footer>
 </div>
@@ -2668,10 +2668,10 @@ select:focus-visible,
     window.eduevidenceCharts = window.eduevidenceCharts || {};
     window.eduevidenceCharts[containerId] = chart;
   }
-  mountChart('chart-outcome-zh', {"chart_id": "outcome-evidence-overview", "purpose": "interactive_analysis", "engine": "echarts", "chart_type": "diverging_bar", "semantic_basis": "effect_direction", "title": "结果证据概览", "option": {"tooltip": {"trigger": "axis"}, "legend": {"data": ["正向效应", "负向效应", "零效应"]}, "grid": [{"left": 150, "right": 40, "top": 30, "height": "52%"}, {"left": 150, "right": 40, "top": "70%", "height": "18%"}], "xAxis": [{"type": "value", "gridIndex": 0, "minInterval": 1}, {"type": "value", "gridIndex": 1, "min": 0, "max": 2, "minInterval": 1}], "yAxis": [{"type": "category", "data": ["Policy Effectiveness", "Implementation Risk"], "inverse": true, "gridIndex": 0}, {"type": "category", "data": ["Policy Effectiveness", "Implementation Risk"], "inverse": true, "gridIndex": 1, "show": false}], "series": [{"name": "正向效应", "type": "bar", "data": [2, 0], "itemStyle": {"color": "#5E8A6A"}, "xAxisIndex": 0, "yAxisIndex": 0, "lane": "main"}, {"name": "负向效应", "type": "bar", "data": [0, -2], "itemStyle": {"color": "#A85B53"}, "xAxisIndex": 0, "yAxisIndex": 0, "lane": "main"}, {"name": "零效应", "type": "bar", "data": [0, 0], "itemStyle": {"color": "#C99A4A"}, "xAxisIndex": 1, "yAxisIndex": 1, "barWidth": 6, "lane": "neutral"}]}, "summary_text": "各 Outcome 的正向 / 负向 / 零效应证据条数对比。这里展示的是 effect_direction，不是“这条证据是否支持某个主张”。", "integrity": {"numbers_match_result": "NOT_CHECKED", "no_axis_distortion": "NOT_CHECKED", "no_false_precision": "NOT_CHECKED", "colorblind_safe": "NOT_CHECKED"}});
+  mountChart('chart-outcome-zh', {"chart_id": "outcome-evidence-overview", "purpose": "interactive_analysis", "engine": "echarts", "chart_type": "diverging_bar", "semantic_basis": "effect_direction", "title": "结果证据概览", "option": {"tooltip": {"trigger": "axis"}, "legend": {"data": ["正向效应", "负向效应", "零效应"]}, "grid": [{"left": 150, "right": 40, "top": 30, "height": "52%"}, {"left": 150, "right": 40, "top": "70%", "height": "18%"}], "xAxis": [{"type": "value", "gridIndex": 0, "minInterval": 1}, {"type": "value", "gridIndex": 1, "min": 0, "max": 2, "minInterval": 1}], "yAxis": [{"type": "category", "data": ["政策有效性", "实施风险"], "inverse": true, "gridIndex": 0}, {"type": "category", "data": ["政策有效性", "实施风险"], "inverse": true, "gridIndex": 1, "show": false}], "series": [{"name": "正向效应", "type": "bar", "data": [2, 0], "itemStyle": {"color": "#5E8A6A"}, "xAxisIndex": 0, "yAxisIndex": 0, "lane": "main"}, {"name": "负向效应", "type": "bar", "data": [0, -2], "itemStyle": {"color": "#A85B53"}, "xAxisIndex": 0, "yAxisIndex": 0, "lane": "main"}, {"name": "零效应", "type": "bar", "data": [0, 0], "itemStyle": {"color": "#C99A4A"}, "xAxisIndex": 1, "yAxisIndex": 1, "barWidth": 6, "lane": "neutral"}]}, "summary_text": "各 Outcome 的正向 / 负向 / 零效应证据条数对比。这里展示的是 effect_direction，不是“这条证据是否支持某个主张”。", "integrity": {"numbers_match_result": "NOT_CHECKED", "no_axis_distortion": "NOT_CHECKED", "no_false_precision": "NOT_CHECKED", "colorblind_safe": "NOT_CHECKED"}});
   mountChart('chart-trace-zh', {"chart_id": "claim-evidence-trace", "purpose": "interactive_analysis", "engine": "echarts", "chart_type": "graph", "title": "主张-证据追溯", "option": {"tooltip": {"trigger": "item"}, "legend": {"data": ["Decision", "Claim", "Evidence", "Source"]}, "series": [{"type": "graph", "layout": "force", "roam": true, "draggable": true, "categories": [{"name": "Decision"}, {"name": "Claim"}, {"name": "Evidence"}, {"name": "Source"}], "data": [{"id": "decision", "name": "PILOT", "category": 0}, {"id": "claim-0", "name": "在有边界的客服场景中，AI 辅助可缩短会话处理时间；吞吐量是另一项指标。", "category": 1}, {"id": "E-001", "name": "E-001", "category": 2}, {"id": "S-001", "name": "S-001", "category": 3}, {"id": "claim-1", "name": "ChatGPT 缩短了短篇职业写作任务用时；对客服而言属于间接证据。", "category": 1}, {"id": "E-002", "name": "E-002", "category": 2}, {"id": "S-002", "name": "S-002", "category": 3}, {"id": "claim-2", "name": "在超出能力边界的任务中 AI 可能降低正确率；咨询实验对客服属于间接证据。", "category": 1}, {"id": "E-003", "name": "E-003", "category": 2}, {"id": "S-003", "name": "S-003", "category": 3}, {"id": "claim-3", "name": "须单独监测资深高技能客服的质量，不能假定所有岗位群体均受益。", "category": 1}, {"id": "E-004", "name": "E-004", "category": 2}], "links": [{"source": "decision", "target": "claim-0"}, {"source": "claim-0", "target": "E-001", "label": "support", "lineStyle": {"color": "#5E8A6A"}}, {"source": "E-001", "target": "S-001"}, {"source": "decision", "target": "claim-1"}, {"source": "claim-1", "target": "E-002", "label": "support", "lineStyle": {"color": "#5E8A6A"}}, {"source": "E-002", "target": "S-002"}, {"source": "decision", "target": "claim-2"}, {"source": "claim-2", "target": "E-003", "label": "support", "lineStyle": {"color": "#5E8A6A"}}, {"source": "E-003", "target": "S-003"}, {"source": "decision", "target": "claim-3"}, {"source": "claim-3", "target": "E-004", "label": "support", "lineStyle": {"color": "#5E8A6A"}}, {"source": "E-004", "target": "S-001"}], "label": {"show": true, "position": "right", "fontSize": 9}, "lineStyle": {"curveness": 0.15}}]}, "summary_text": "决策→结论→证据→来源 的可追溯图谱；点击节点可追踪支持/反驳路径。", "integrity": {"numbers_match_result": "NOT_CHECKED", "no_axis_distortion": "NOT_CHECKED", "no_false_precision": "NOT_CHECKED", "colorblind_safe": "NOT_CHECKED"}});
   mountChart('chart-benchmark-zh', {"chart_id": "benchmark-panel", "purpose": "interactive_analysis", "engine": "echarts", "chart_type": "composite", "title": "基准测试：B0-B4", "option": {"tooltip": {"trigger": "axis"}, "legend": {"data": ["Citation Support", "Unsupported Rate", "Contradiction"]}, "grid": {"left": 60, "right": 40}, "xAxis": {"type": "category", "data": []}, "yAxis": {"type": "value", "max": 1, "min": 0}, "series": [{"name": "Citation Support", "type": "bar", "data": []}, {"name": "Unsupported Rate", "type": "bar", "data": []}, {"name": "Contradiction", "type": "bar", "data": []}]}, "cost_vs_quality": {"chart_id": "benchmark-quality-cost", "purpose": "interactive_analysis", "engine": "echarts", "chart_type": "scatter", "title": "质量 vs 成本", "option": {"tooltip": {"trigger": "item"}, "xAxis": {"type": "value", "name": "cost (USD)", "min": 0}, "yAxis": {"type": "value", "name": "citation support", "min": 0, "max": 1}, "series": [{"type": "scatter", "data": []}]}}, "summary_text": "B0-B4 基线在引用支持/无支撑率/反方发现上的对比及质量-成本散点。", "integrity": {"numbers_match_result": "NOT_CHECKED", "no_axis_distortion": "NOT_CHECKED", "no_false_precision": "NOT_CHECKED", "colorblind_safe": "NOT_CHECKED"}});
-  mountChart('chart-outcome-en', {"chart_id": "outcome-evidence-overview", "purpose": "interactive_analysis", "engine": "echarts", "chart_type": "diverging_bar", "semantic_basis": "effect_direction", "title": "Outcome Evidence Overview", "option": {"tooltip": {"trigger": "axis"}, "legend": {"data": ["Positive effect", "Negative effect", "Null effect"]}, "grid": [{"left": 150, "right": 40, "top": 30, "height": "52%"}, {"left": 150, "right": 40, "top": "70%", "height": "18%"}], "xAxis": [{"type": "value", "gridIndex": 0, "minInterval": 1}, {"type": "value", "gridIndex": 1, "min": 0, "max": 2, "minInterval": 1}], "yAxis": [{"type": "category", "data": ["Policy Effectiveness", "Implementation Risk"], "inverse": true, "gridIndex": 0}, {"type": "category", "data": ["Policy Effectiveness", "Implementation Risk"], "inverse": true, "gridIndex": 1, "show": false}], "series": [{"name": "Positive effect", "type": "bar", "data": [2, 0], "itemStyle": {"color": "#5E8A6A"}, "xAxisIndex": 0, "yAxisIndex": 0, "lane": "main"}, {"name": "Negative effect", "type": "bar", "data": [0, -2], "itemStyle": {"color": "#A85B53"}, "xAxisIndex": 0, "yAxisIndex": 0, "lane": "main"}, {"name": "Null effect", "type": "bar", "data": [0, 0], "itemStyle": {"color": "#C99A4A"}, "xAxisIndex": 1, "yAxisIndex": 1, "barWidth": 6, "lane": "neutral"}]}, "summary_text": "Positive / negative / null effect-direction evidence counts per outcome. This visual encodes effect_direction, not whether evidence supports a claim.", "integrity": {"numbers_match_result": "NOT_CHECKED", "no_axis_distortion": "NOT_CHECKED", "no_false_precision": "NOT_CHECKED", "colorblind_safe": "NOT_CHECKED"}});
+  mountChart('chart-outcome-en', {"chart_id": "outcome-evidence-overview", "purpose": "interactive_analysis", "engine": "echarts", "chart_type": "diverging_bar", "semantic_basis": "effect_direction", "title": "Outcome Evidence Overview", "option": {"tooltip": {"trigger": "axis"}, "legend": {"data": ["Positive effect", "Negative effect", "Null effect"]}, "grid": [{"left": 150, "right": 40, "top": 30, "height": "52%"}, {"left": 150, "right": 40, "top": "70%", "height": "18%"}], "xAxis": [{"type": "value", "gridIndex": 0, "minInterval": 1}, {"type": "value", "gridIndex": 1, "min": 0, "max": 2, "minInterval": 1}], "yAxis": [{"type": "category", "data": ["Policy effectiveness", "Implementation risk"], "inverse": true, "gridIndex": 0}, {"type": "category", "data": ["Policy effectiveness", "Implementation risk"], "inverse": true, "gridIndex": 1, "show": false}], "series": [{"name": "Positive effect", "type": "bar", "data": [2, 0], "itemStyle": {"color": "#5E8A6A"}, "xAxisIndex": 0, "yAxisIndex": 0, "lane": "main"}, {"name": "Negative effect", "type": "bar", "data": [0, -2], "itemStyle": {"color": "#A85B53"}, "xAxisIndex": 0, "yAxisIndex": 0, "lane": "main"}, {"name": "Null effect", "type": "bar", "data": [0, 0], "itemStyle": {"color": "#C99A4A"}, "xAxisIndex": 1, "yAxisIndex": 1, "barWidth": 6, "lane": "neutral"}]}, "summary_text": "Positive / negative / null effect-direction evidence counts per outcome. This visual encodes effect_direction, not whether evidence supports a claim.", "integrity": {"numbers_match_result": "NOT_CHECKED", "no_axis_distortion": "NOT_CHECKED", "no_false_precision": "NOT_CHECKED", "colorblind_safe": "NOT_CHECKED"}});
   mountChart('chart-trace-en', {"chart_id": "claim-evidence-trace", "purpose": "interactive_analysis", "engine": "echarts", "chart_type": "graph", "title": "Claim-Evidence Trace", "option": {"tooltip": {"trigger": "item"}, "legend": {"data": ["Decision", "Claim", "Evidence", "Source"]}, "series": [{"type": "graph", "layout": "force", "roam": true, "draggable": true, "categories": [{"name": "Decision"}, {"name": "Claim"}, {"name": "Evidence"}, {"name": "Source"}], "data": [{"id": "decision", "name": "PILOT", "category": 0}, {"id": "claim-0", "name": "AI assistance can shorten customer chat ", "category": 1}, {"id": "E-001", "name": "E-001", "category": 2}, {"id": "S-001", "name": "S-001", "category": 3}, {"id": "claim-1", "name": "ChatGPT reduced time on short profession", "category": 1}, {"id": "E-002", "name": "E-002", "category": 2}, {"id": "S-002", "name": "S-002", "category": 3}, {"id": "claim-2", "name": "AI can reduce correctness on tasks outsi", "category": 1}, {"id": "E-003", "name": "E-003", "category": 2}, {"id": "S-003", "name": "S-003", "category": 3}, {"id": "claim-3", "name": "Experienced, high-skill support staff ne", "category": 1}, {"id": "E-004", "name": "E-004", "category": 2}], "links": [{"source": "decision", "target": "claim-0"}, {"source": "claim-0", "target": "E-001", "label": "support", "lineStyle": {"color": "#5E8A6A"}}, {"source": "E-001", "target": "S-001"}, {"source": "decision", "target": "claim-1"}, {"source": "claim-1", "target": "E-002", "label": "support", "lineStyle": {"color": "#5E8A6A"}}, {"source": "E-002", "target": "S-002"}, {"source": "decision", "target": "claim-2"}, {"source": "claim-2", "target": "E-003", "label": "support", "lineStyle": {"color": "#5E8A6A"}}, {"source": "E-003", "target": "S-003"}, {"source": "decision", "target": "claim-3"}, {"source": "claim-3", "target": "E-004", "label": "support", "lineStyle": {"color": "#5E8A6A"}}, {"source": "E-004", "target": "S-001"}], "label": {"show": true, "position": "right", "fontSize": 9}, "lineStyle": {"curveness": 0.15}}]}, "summary_text": "Traceable graph Decision → Claim → Evidence → Source; click nodes to follow support/contradict paths.", "integrity": {"numbers_match_result": "NOT_CHECKED", "no_axis_distortion": "NOT_CHECKED", "no_false_precision": "NOT_CHECKED", "colorblind_safe": "NOT_CHECKED"}});
   mountChart('chart-benchmark-en', {"chart_id": "benchmark-panel", "purpose": "interactive_analysis", "engine": "echarts", "chart_type": "composite", "title": "Benchmark: B0-B4", "option": {"tooltip": {"trigger": "axis"}, "legend": {"data": ["Citation Support", "Unsupported Rate", "Contradiction"]}, "grid": {"left": 60, "right": 40}, "xAxis": {"type": "category", "data": []}, "yAxis": {"type": "value", "max": 1, "min": 0}, "series": [{"name": "Citation Support", "type": "bar", "data": []}, {"name": "Unsupported Rate", "type": "bar", "data": []}, {"name": "Contradiction", "type": "bar", "data": []}]}, "cost_vs_quality": {"chart_id": "benchmark-quality-cost", "purpose": "interactive_analysis", "engine": "echarts", "chart_type": "scatter", "title": "Quality vs Cost", "option": {"tooltip": {"trigger": "item"}, "xAxis": {"type": "value", "name": "cost (USD)", "min": 0}, "yAxis": {"type": "value", "name": "citation support", "min": 0, "max": 1}, "series": [{"type": "scatter", "data": []}]}}, "summary_text": "B0-B4 baselines compared on citation support, unsupported rate and contradiction discovery, plus the quality-cost scatter.", "integrity": {"numbers_match_result": "NOT_CHECKED", "no_axis_distortion": "NOT_CHECKED", "no_false_precision": "NOT_CHECKED", "colorblind_safe": "NOT_CHECKED"}});
 })();