跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
数据与分析

High Stakes Analytics Decision Lab

A platform-neutral analytical skill that profiles messy data, selects adaptive methods, and produces source-backed visual reports for high-stakes decisions.

limingrui679-design · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add limingrui679-design/high-stakes-analytics-decision-lab

技能说明


name: high-stakes-analytics-decision-lab description: Build or review source-backed descriptive, diagnostic, predictive, and prescriptive analysis for consequential decisions. Use when an agent must profile and safely prepare uploaded data, turn a real dataset or research question into a reproducible study, investigate drivers without overstating causality, validate a model, compare feasible actions under dependent uncertainty and tail risk, trace every parameter to evidence and approval, or produce an answer-first analytical report across health, business, finance, policy, engineering, operations, behavioral science, AI, or planning.

High-Stakes Analytics & Decision Lab

Turn a real question into a defensible path from source evidence to analysis and, only when justified, bounded action. Keep observation, diagnosis, prediction, causal evidence, value judgments, and recommendation visibly separate.

Start here

Run the environment audit before an executable workflow:

python3 scripts/hsadl.py doctor

For a safe end-to-end setup example:

python3 scripts/hsadl.py demo --output-dir /absolute/path/to/demo

The demo is a synthetic engineering fixture. Never cite its values as empirical evidence.

Route before choosing a method

RoutePrimary questionValid endpoint
DescriptiveWhat is happening?Baseline report or evidence request
DiagnosticWhy might it be happening?Explanations to test with a visible causal boundary
PredictiveWhat is likely next?Validated prediction, negative validation, or do_not_deploy
PrescriptiveWhat should be done, if justified?Bounded action, pilot, diligence, evidence request, or no recommendation

When only a question is available, generate a blueprint instead of inventing results:

python3 scripts/hsadl.py route "How should limited review capacity be allocated?" \
  --scope full --output-dir /absolute/path/to/blueprint

Read references/analytics-triad.md and references/method-routing.md when a request is ambiguous or spans several routes.

Evidence-gated workflow

  1. Define the contract. State the decision or research question, population, analytical unit, target quantity, horizon, intended use, stakeholders, and claim boundary.
  2. Establish lineage. Prefer official, academic, or otherwise authoritative sources. Record publisher, version, access date, license, redistribution rule, grain, exclusions, file paths, and SHA-256 hashes.
  3. Gate the data. Preserve every supplied source unchanged. Profile grain, keys, schema, completeness, type and domain validity, time reliability, privacy signals, and target leakage before calculating a result.
  4. Build the baseline. Define denominators, coverage, missingness, trends, segments, and comparability before diagnosis, prediction, or action.
  5. Add only justified modules. Select methods from the question, estimand, data-generating structure, and decision; never from column availability alone.
  6. Validate and challenge. Use a defensible holdout or identification strategy, baseline comparisons, calibration or uncertainty, subgroup or distribution checks, dependence-aware stress, sensitivity, and reversal conditions as applicable.
  7. Communicate the strongest supported claim—no stronger. The Evidence Intelligence Report is primary. Add a Decision Intelligence Brief only when a real decision, feasible alternatives, and sufficient evidence exist.

Read references/real-evidence-workflow.md, references/data-quality-gate.md, and references/methodology.md for the full contract.

Start from a real dataset

Create a reviewable workspace in one command:

python3 scripts/hsadl.py start /absolute/path/to/input.csv \
  --question "Which groups are likely to need support next month?" \
  --output-dir /absolute/path/to/workspace

The initializer copies and hash-checks the source, drafts or accepts a data contract, profiles readiness, routes the question, and records unresolved decisions. It must not clean data, fit a model, or generate a recommendation.

The gate returns exactly one of:

  • ready;
  • ready_with_documented_limitations;
  • needs_user_confirmation;
  • blocked.

Continue only when the gate permits the intended route. Run only safe_auto normalization without approval. Deletion, column removal, imputation, outlier treatment, category merging, unit or timezone conversion, target correction, and grain changes require approval by the exact action ID. Fail closed if the source hash, reviewed action, approval, or raw/processed binding changes.

Direct gate and preparation commands:

python3 scripts/hsadl.py profile input.csv \
  --contract data-contract.json --output-dir readiness

python3 scripts/hsadl.py prepare input.csv \
  --quality-report readiness/data-quality-report.json \
  --cleaning-plan readiness/cleaning-plan.json \
  --approve clean-003 --output-dir prepared

Select an executable module

NeedCommandRequired boundary
Two-group binary, continuous, or time-to-event evidencehsadl.py evidenceMatch the estimand and study design
Held-out scores, calibration, subgroup error, or drifthsadl.py predictPrediction is not intervention effect
Small discrete allocation with constraints and scenarioshsadl.py allocateInputs and objectives are not empirical facts by default
Multi-criterion decision under dependent uncertaintyhsadl.py validate then hsadl.py runRequire owner, alternatives, constraints, provenance, approval, tails, sensitivity, and affected groups

Read references/method-modules.md for command contracts and references/advanced-method-boundaries.md before survival, repeated-measures, financial-risk, spatial, or responsible-AI work.

Decision layer

Do not begin simulation while the owner, decision, alternatives, horizon,

阅读完整来源 ↗