Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when designing or writing heterogeneity / conditionality analysis for a 《中国行政管理》(Chinese Public Administration, CPA) manuscript — variation across regions, administrative levels, and governance contexts. 当主结果只笼统适用、只切了一个维度、或审稿人要求差异化分析时使用。Enforces theory-guided cuts and significance-of-difference discipline for this Chinese-Public-Administration-Society journal.
.claude/skills/brycewang-stanford-cpa-heterogeneity/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 24% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -3% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 51% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 50% | 0% |
| case-22 | ✗→✓ | ▲ Improved | 74% | 0% |
按《中国行政管理》读者期待度排序（公共管理语境）：
至少切 2–3 个维度；定量每个维度至少 2 个子样本对比，定性按维度做系统的跨案例比较。
本文进一步检验[维度，如行政层级]的异质性。
理论上，[原因，挂机制]……因此预期效应在[组 1]更强。
实证上，本文将样本按[维度]分为[组 1]和[组 2]，分别估计（结果见表 X 第 (1)(2) 列）。
结果显示，[组 1]的效应为 X，[组 2]为 Y，系数差异在[显著水平]上显著。
这一发现与本文的[机制]一致——[解释]。本文比较[案例 A / B / C]在[结果]上的差异。
通过对[条件 1、条件 2、条件 3]的系统比较，发现导向[成功 / 失败]的并非单一条件，而是[条件组合]。
这表明在[治理情境]下，[结果]具有[多重并发因果 / 殊途同归]特征。异质性 ≈ 机制的反向验证：
理想结构：机制（某条件激活该过程） 与 异质性（该条件强的样本效应强） 相互印证。
《中国行政管理》（中国行政管理学会主办，CSSCI 权威公共管理顶级期刊）审稿人对异质性的核心期待，是看它能否照见中国治理结构本身——压力型体制、央地分工、属地管理、层级治理逻辑。把异质性切在能体现治理结构的维度上，比切人口学变量更被本刊认可。
| 切分倾向 | 本刊更认可 | 易被判"凑数" | |---------|----------|------------| | 治理结构 | 行政层级、属地加码强度、条块关系 | 仅按性别 / 年龄切个体特征 | | 治理情境 | 数字政府发展水平、财政能力、市场化—法治化程度 | 仅东中西三分而无机制解释 | | 政策窗口 | 运动式治理期 vs. 常态化、机构改革前后 | 随机切年份无理论指引 |
> 数字政府、应急管理、基层治理类稿件，本刊尤其期待看到"同一政策在不同治理能力地区效果不同"的条件性结论。维度命名是否被认可，以编辑部最新稿约与近期同栏目论文为准。
以一篇虚构的"网格化治理对社区矛盾化解的影响"评估稿件演示。下列数字为示意，仅说明切分与差异检验流程。
主结果（示意）：网格化使社区矛盾化解率上升约 8 个百分点。
切分维度1（行政层级 / 属地加码）：
强属地加码区 效应 +11pp ，弱属地加码区 +4pp ，
系数差异 Chow 检验在 5% 上显著 → 与"层层加码激活基层响应"机制一致。
切分维度2（数字政府水平）：
高数字化社区 +12pp ，低数字化社区 +5pp ，差异显著
→ 印证"技术赋能放大网格响应"这一机制。
结论：两条异质性均回扣机制，而非孤立罗列。走查结论：维度紧扣治理结构与数字治理情境，每个维度都报告系数差异检验并回扣机制——满足本刊"理论指引 + 显著性纪律"双重要求。
cpa-mechanism，让"激活该机制的条件"与"该条件强的子样本效应强"对齐。【异质性维度】X 个（层级 / 情境 / 结构 / ……）
【差异检验 / 比较方式】系数差异检验 / 跨案例条件比较 / QCA
【与机制一致性】是 / 否
【最小子样本 / 案例数】X
【下一步】cpa-tables-figures| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 30,972 | 26,126 | -16% | 1 | 1 | 0% | 3,829 | 4,753 | +24% | 0 | 0 | — |
case-02 | fail→pass | 37,148 | 24,651 | -34% | 1 | 1 | 0% | 4,833 | 4,704 | -3% | 0 | 0 | — |
case-03 | fail→pass | 22,792 | 22,912 | +1% | 1 | 1 | 0% | 2,784 | 4,216 | +51% | 0 | 0 | — |
case-04 | pass→pass | 24,862 | 33,989 | +37% | 1 | 1 | 0% | 3,034 | 4,796 | +58% | 0 | 0 | — |
case-05 | fail→fail | 27,907 | 17,869 | -36% | 1 | 1 | 0% | 2,952 | 4,249 | +44% | 0 | 0 | — |
case-06 | fail→pass | 30,272 | 27,515 | -9% | 1 | 1 | 0% | 3,280 | 4,911 | +50% | 0 | 0 | — |
case-20 | pass→pass | 16,763 | 19,747 | +18% | 1 | 1 | 0% | 2,099 | 4,422 | +111% | 0 | 0 | — |
case-21 | fail→fail | 23,787 | 19,092 | -20% | 1 | 1 | 0% | 3,511 | 4,476 | +27% | 0 | 0 | — |
case-22 | fail→pass | 21,348 | 27,033 | +27% | 1 | 1 | 0% | 2,998 | 5,223 | +74% | 0 | 0 | — |
case-23 | fail→fail | 22,712 | 19,641 | -14% | 1 | 1 | 0% | 2,908 | 4,367 | +50% | 0 | 0 | — |
case-07 | pass→pass | 26,620 | 25,250 | -5% | 1 | 1 | 0% | 3,110 | 4,673 | +50% | 0 | 0 | — |
case-08 | pass→pass | 20,362 | 24,310 | +19% | 1 | 1 | 0% | 2,548 | 4,287 | +68% | 0 | 0 | — |
case-09 | pass→pass | 21,082 | 28,097 | +33% | 1 | 1 | 0% | 2,850 | 4,688 | +64% | 0 | 0 | — |
case-10 | fail→pass | 16,937 | 13,153 | -22% | 1 | 1 | 0% | 2,348 | 2,808 | +20% | 0 | 0 | — |
case-11 | pass→pass | 35,334 | 23,226 | -34% | 1 | 1 | 0% | 3,647 | 4,575 | +25% | 0 | 0 | — |
case-12 | pass→pass | 22,028 | 28,441 | +29% | 1 | 1 | 0% | 2,946 | 4,494 | +53% | 0 | 0 | — |
case-13 | pass→pass | 20,398 | 18,597 | -9% | 1 | 1 | 0% | 3,064 | 3,654 | +19% | 0 | 0 | — |
case-14 | fail→fail | 21,066 | 18,584 | -12% | 1 | 1 | 0% | 3,472 | 3,583 | +3% | 0 | 0 | — |
case-15 | pass→pass | 21,508 | 20,985 | -2% | 1 | 1 | 0% | 3,076 | 4,213 | +37% | 0 | 0 | — |
case-16 | pass→pass | 28,100 | 21,875 | -22% | 1 | 1 | 0% | 3,358 | 4,676 | +39% | 0 | 0 | — |
case-17 | pass→pass | 25,682 | 20,789 | -19% | 1 | 1 | 0% | 3,225 | 4,538 | +41% | 0 | 0 | — |
case-18 | pass→pass | 18,750 | 20,328 | +8% | 1 | 1 | 0% | 3,206 | 5,162 | +61% | 0 | 0 | — |
case-19 | pass→pass | 24,370 | 25,922 | +6% | 1 | 1 | 0% | 3,262 | 5,026 | +54% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of +26 percentage points is the difference between those two pass rates over the 23 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.