experiment-execution — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited experiment-execution (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
定时执行(每日/每4小时)
├── 数据同步检查
├── 样本量进度
├── 主指标趋势
├── 护栏指标检查
├── 统计显著性检查
└── 异常检测触发条件:达到终止条件
├── 锁定数据
├── 统计分析
├── 下钻分析
├── 生成结论
├── 生成报告
└── 输出建议| 输入项 | 类型 | 必填 | 来源 | 说明 |
|---|---|---|---|---|
| 实验设计文档 | object | 是 | output/pm-metrics-ops/experiment-design/experiment_design.json | experiment-design 输出的实验方案 |
| 实验数据 | object | 是 | 用户提供 | 分组数据、指标数据、护栏指标数据 |
| 终止条件 | object | 是 | output/pm-metrics-ops/experiment-design/experiment_design.json | 样本量目标、运行时长、最小可检测效应 |
| 产品背景 | text | 否 | 用户输入 | 产品阶段、业务目标、历史实验 |
统计检验、业务意义评估、多维下钻、新奇效应检测
#### 1.1 统计显著性检验
##### 检验方法选择
| 指标类型 | 检验方法 | 说明 |
|---|---|---|
| 比例(转化率) | Z-test / 卡方检验 | 二项分布 |
| 均值(收入) | T-test / Mann-Whitney | 正态/非正态 |
| 分布(时长) | KS检验 | 分布差异 |
| 多指标 | FDR校正 | 多重检验 |
##### 输出格式
statistical_test:
method: "two_sample_proportion_z_test"
results:
control:
sample_size: 12450
rate: 0.352
standard_error: 0.0043
treatment:
sample_size: 12380
rate: 0.381
standard_error: 0.0044
test_statistic: 4.82
p_value: 0.0000014
confidence_interval:
lower: 0.018
upper: 0.040
confidence_level: 0.95
interpretation:
is_significant: true
significance_level: 0.05
conclusion: "实验组显著优于对照组"#### 1.2 实际业务意义评估
统计显著 ≠ 业务有效
practical_significance:
absolute_lift: 0.029
relative_lift: 0.082
threshold:
minimum_meaningful_lift: 0.02
assessment:
is_practically_significant: true
business_verdict: "值得全量发布"
reasoning: "提升8.2%超过业务阈值5%,预计年收入增加120万"#### 1.3 多维下钻分析
##### 异质性效应检测
heterogeneous_effects:
summary: "发现显著的平台差异"
dimension_analysis:
platform:
ios:
lift: 0.052
p_value: 0.001
significant: true
android:
lift: 0.018
p_value: 0.089
significant: false
conclusion: "iOS用户效果显著,Android不显著"
user_segment:
new_users:
lift: 0.041
significant: true
returning_users:
lift: 0.015
significant: false
conclusion: "主要对新用户有效"
traffic_source:
organic:
lift: 0.045
significant: true
paid:
lift: 0.022
significant: false
conclusion: "对自然流量更有效"
recommendations:
- "考虑仅在iOS全量"
- "优化Android版本的实现"
- "针对新用户定向推广"#### 1.4 新奇效应检测
检测用户初期行为异常:
novelty_check:
enabled: true
indicators:
daily_trend:
day_1: 0.15
day_3: 0.09
day_7: 0.08
day_14: 0.082
assessment:
is_novelty_effect: false
trend_stable: true
conclusion: "效应稳定,无新奇效应"
actions:
if_novelty: "延长实验周期2周"
if_stable: "可进入决策流程"#### 1.5 决策建议生成
decision_recommendation:
overall_assessment:
statistical_significance: true
practical_significance: true
guardrail_metrics_safe: true
no_heterogeneous_risks: true
novelty_effect_resolved: true
conclusion: "positive"
primary_metric:
name: "registration_completion_rate"
control:
value: 0.352
lower_ci: 0.344
upper_ci: 0.360
treatment:
value: 0.381
lower_ci: 0.373
upper_ci: 0.389
lift:
absolute: 0.029
relative: 0.082
confidence_interval: [0.018, 0.040]
statistics:
p_value: 0.0000014
statistically_significant: true
practically_significant: true
guardrail_metrics:
- name: "d7_retention_rate"
control: 0.42
treatment: 0.418
change: -0.47%
safe: true
verdict: "无显著影响"
- name: "daily_active_users"
control: 1000000
treatment: 1001500
change: +0.15%
safe: true
verdict: "无显著影响"
- name: "app_crash_rate"
control: 0.002
treatment: 0.0021
change: +5%
safe: true
verdict: "无显著影响"
heterogeneous_effects:
summary: "iOS效果显著(+5.2%),Android不显著(+1.8%)"
recommendations:
- "考虑分平台发布策略"
- "Android版本需要进一步优化"
novelty_check:
detected: false
trend: "stable"
recommendation:
action: "全量发布"
confidence: "high"
reasoning:
- "主指标提升8.2%,统计显著"
- "实际业务意义显著"
- "护栏指标安全"
- "效应稳定,无新奇效应"
risks:
- "Android效果不确定,需后续监控"
next_steps:
- "全量发布到iOS和Android"
- "发布后2周监控关键指标"
- "如果Android表现持续不佳,考虑回滚"实验概述、统计结论、效果分析、行动建议
#### 2.1 实验概述组装
从实验设计方案提取核心要素:
#### 2.2 统计结论提炼
从实验执行结果提炼统计结论:
#### 2.3 效果深度分析
对核心效果进行多维度分析:
#### 2.4 行动建议生成
基于统计结论和效果分析,生成分层行动建议:
| 结论类型 | 建议等级 | 行动 |
|---|---|---|
| 核心指标显著正向 + 护栏安全 | 🟢 强烈推荐全量 | 全量发布 + 监控计划 |
| 核心指标显著正向 + 护栏有风险 | 🟡 条件推荐 | 分阶段全量 + 护栏专项优化 |
| 核心指标不显著 | 🔵 需要更多信息 | 延长周期/扩大样本/调整指标 |
| 核心指标显著负向 | 🔴 建议终止 | 终止实验 + 根因分析 |
| 异质性效应显著 | 🟠 分群策略 | 分群差异化发布 |
#### 2.5 报告组装
将以上内容组装为完整报告。
存储路径:output/pm-metrics-ops/experiment-execution/
输出文件:
| 文件 | 路径 | 说明 |
|---|---|---|
| 实验结果数据 | output/pm-metrics-ops/experiment-execution/experiment_result.json | 机器可消费的实验结果数据 |
| A/B测试报告 | output/pm-metrics-ops/experiment-execution/experiment-report.md | 人类可读的完整报告 |
| 结构化报告数据 | output/pm-metrics-ops/experiment-execution/experiment-report.json | 机器可消费的结构化报告数据 |
输出Schema:
{
"type": "object",
"required": ["experiment_id", "conclusion", "primary_metric", "summary", "action_recommendation"],
"properties": {
"experiment_id": {"type": "string", "description": "实验ID"},
"analyzed_at": {"type": "string", "description": "分析时间"},
"experiment_info": {"type": "object", "description": "实验信息,包含名称、周期和样本量"},
"conclusion": {"type": "string", "description": "实验结论:positive/negative/neutral/inconclusive"},
"primary_metric": {"type": "object", "description": "主指标结果,包含对照组/实验组数据和统计检验"},
"guardrail_metrics": {"type": "object", "description": "护栏指标结果,包含各指标变化和安全性判断"},
"heterogeneous_effects": {"type": "object", "description": "异质性效应,按平台/用户类型分群分析"},
"novelty_check": {"type": "object", "description": "新奇效应检测"},
"experiment_name": {"type": "string", "description": "实验名称"},
"report_date": {"type": "string", "description": "报告日期"},
"summary": {"type": "object", "description": "统计结论摘要,包含结论、推荐和主指标结果"},
"novelty_effect": {"type": "object", "description": "新奇效应评估"},
"action_recommendation": {"type": "object", "description": "行动建议,包含决策、理由、风险和后续实验"}
}
}ab_test_result:
experiment_id: "exp_20240115_simplified_register"
analyzed_at: "2024-01-22T10:00:00Z"
experiment_info:
name: "简化注册流程实验"
start_date: "2024-01-15"
end_date: "2024-01-21"
duration_days: 7
total_sample: 24830
conclusion: "positive"
primary_metric:
name: "registration_completion_rate"
control:
value: 0.352
sample_size: 12450
confidence_interval: [0.344, 0.360]
treatment:
value: 0.381
sample_size: 12380
confidence_interval: [0.373, 0.389]
lift:
absolute: 0.029
relative: 0.082
confidence_interval: [0.018, 0.040]
statistics:
p_value: 0.0000014
test_method: "two_sample_z_test"
statistically_significant: true
practically_significant: true
guardrail_metrics:
d7_retention_rate:
control: 0.42
treatment: 0.418
change: -0.47%
safe: true
daily_active_users:
control: 1000000
treatment: 1001500
change: +0.15%
safe: true
app_crash_rate:
control: 0.002
treatment: 0.0021
change: +5%
safe: true
heterogeneous_effects:
platform:
ios: { lift: 0.052, significant: true }
android: { lift: 0.018, significant: false }
user_type:
new_users: { lift: 0.041, significant: true }
returning_users: { lift: 0.015, significant: false }
novelty_check:
detected: false
trend_stable: true
decision_recommendation:
action: "full_release"
confidence: "high"
reasoning:
- "主指标提升8.2%"
- "护栏指标安全"
- "无新奇效应"# A/B测试报告:{实验名称}
## 1. 实验概述
- 实验ID / 运行周期 / 样本量
- 假设陈述(H₀ / H₁)
- 指标体系(核心 / 护栏 / 辅助)
- 分流方案
## 2. 统计结论
- 核心指标:效应量 [CI] (p=xxx)
- 护栏指标:✅/⚠️/❌ 逐项检查
- 样本量验证:达标/未达标
- 综合判定:显著正向 / 不显著 / 显著负向
## 3. 效果分析
- 效应量解读(绝对/相对/业务换算)
- 异质性效应(分群下钻表格)
- 新奇效应评估
- 交互效应检查
## 4. 行动建议
- 推荐行动 + 理由
- 风险提示
- 后续实验建议
## 5. 附录
- 统计方法说明
- 数据质量检查
- 完整指标明细表{
"experiment_id": "",
"experiment_name": "",
"report_date": "",
"summary": {
"conclusion": "significant_positive|not_significant|significant_negative|marginal",
"recommendation": "ship_full|ship_conditional|extend|terminate|segmented",
"primary_metric": {
"name": "",
"control_value": 0,
"treatment_value": 0,
"absolute_lift": 0,
"relative_lift": 0,
"confidence_interval": [0, 0],
"p_value": 0,
"statistical_power": 0
},
"guardrail_status": []
},
"heterogeneous_effects": [],
"novelty_effect": {},
"action_recommendation": {
"decision": "",
"rationale": "",
"risks": [],
"next_experiments": []
}
}| 字段路径 | 类型 | 必填 | 说明 |
|---|---|---|---|
| ab_test_result | object | 是 | 实验结果根对象 |
| ab_test_result.experiment_id | string | 是 | 实验ID |
| ab_test_result.analyzed_at | string | 是 | 分析时间 |
| ab_test_result.conclusion | string | 是 | 实验结论,枚举值:positive/negative/neutral/inconclusive |
| ab_test_result.primary_metric | object | 是 | 主指标结果 |
| ab_test_result.primary_metric.name | string | 是 | 主指标名称 |
| ab_test_result.primary_metric.control.value | number | 是 | 对照组值 |
| ab_test_result.primary_metric.treatment.value | number | 是 | 实验组值 |
| ab_test_result.primary_metric.lift.relative | number | 是 | 相对提升 |
| ab_test_result.primary_metric.statistics.p_value | number | 是 | p值 |
| ab_test_result.primary_metric.statistics.statistically_significant | boolean | 是 | 是否统计显著 |
| ab_test_result.guardrail_metrics | object | 是 | 护栏指标结果 |
| ab_test_result.heterogeneous_effects | object | 否 | 异质性效应 |
| ab_test_result.novelty_check | object | 是 | 新奇效应检测 |
| ab_test_result.novelty_check.detected | boolean | 是 | 是否检测到新奇效应 |
| ab_test_result.decision_recommendation | object | 是 | 决策建议 |
| ab_test_result.decision_recommendation.action | string | 是 | 建议行动,枚举值:full_release/partial_release/no_release/continue_experiment |
| ab_test_result.decision_recommendation.confidence | string | 是 | 置信度 |
| experiment_id | string | 是 | 实验ID(报告) |
| experiment_name | string | 是 | 实验名称(报告) |
| report_date | string | 是 | 报告日期 |
| summary | object | 是 | 统计结论摘要 |
| summary.conclusion | string | 是 | 结论,枚举值:significant_positive/not_significant/significant_negative/marginal |
| summary.recommendation | string | 是 | 建议,枚举值:ship_full/ship_conditional/extend/terminate/segmented |
| summary.primary_metric | object | 是 | 主指标结果 |
| summary.primary_metric.name | string | 是 | 主指标名称 |
| summary.primary_metric.relative_lift | number | 是 | 相对提升 |
| summary.primary_metric.p_value | number | 是 | p值 |
| summary.guardrail_status | array | 是 | 护栏指标状态列表 |
| heterogeneous_effects | array | 否 | 异质性效应分析 |
| novelty_effect | object | 否 | 新奇效应评估 |
| action_recommendation | object | 是 | 行动建议 |
| action_recommendation.decision | string | 是 | 决策 |
| action_recommendation.rationale | string | 是 | 理由 |
| action_recommendation.risks | array | 是 | 风险列表 |
| action_recommendation.next_experiments | array | 否 | 后续实验建议 |
当上游输入发生变更时,本Skill的响应策略:
| 上游变更 | 影响范围 | 响应策略 |
|---|---|---|
| 实验设计变更 | 统计检验参数和终止条件 | 更新统计检验配置,重新评估终止条件 |
| 实验数据更新 | 统计检验和下钻分析 | 重新执行统计检验,更新异质性效应 |
| 终止条件变更 | 实验运行监控 | 更新终止条件,重新评估是否达到终止标准 |
| 产品背景变更 | 行动建议的业务针对性 | 重新评估行动建议,更新风险和后续实验建议 |
当实验结果/报告自身变更时,对下游的通知机制:
| 结果/报告变更类型 | 通知范围 | 通知方式 |
|---|---|---|
| 结论变更 | decision-dace | 标记结论变更,触发DACE Analyze |
| 护栏指标触发告警 | decision-dace | 标记护栏告警,触发洞察转化 |
| 决策建议变更 | decision-dace | 标记建议变更,触发DACE Conclude |
| 行动建议变更 | decision-culture | 标记建议变更,触发报告推送 |
| 条件组合 | 决策 |
|---|---|
| 主指标显著+有意义,护栏安全 | 全量发布 |
| 主指标显著,护栏有问题 | 分析护栏原因,再决策 |
| 主指标不显著 | 继续实验或终止 |
| 有新奇效应 | 延长实验 |
| 异质性显著 | 分群体发布 |
| 缺失范围 | 降级方案 | 输出影响 |
|---|---|---|
| 实验配置缺失 | 无法自动监控,需用户提供实验结果数据 | 无法执行运行中监控 |
| 实验数据缺失 | 用户提供实验结果数据 → 直接分析 | 无法进行趋势分析和新奇效应检测 |
| 实验配置 + 实验数据均缺失 | 用户提供实验结果数据 → 直接分析 | 输出基于用户数据的分析结果,趋势和新奇效应标注"待补充" |
| 无实验设计方案 | 基于执行结果反推实验设计要素,标注"设计信息缺失" | 实验概述章节不完整 |
| 无产品背景 | 聚焦统计结论本身,行动建议标注"需结合业务上下文" | 行动建议可能缺乏业务针对性 |
当上游文件缺失时,需用户提供以下信息以支撑降级生成:
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.