菜单
Use when about to claim that an experiment result, metric validation, or notebook analysis is final, significant, trustworthy, or decision-ready and you need fresh rerun evidence instead of stale outputs
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when a research question spans multiple SQL, pandas, or notebook steps and you need explicit cohorts, windows, metrics, validation checks, outputs, and statistical methods before touching queries or cells
Use when you already have a written analysis plan and want to execute SQL, notebook, or validation tasks in a separate session with review checkpoints and rerun-based evidence between batches
Use when a metric definition, denominator, CUPED covariate, linearization choice, invariant, outlier treatment, missingness pattern, or leakage risk may distort an experiment or validation readout
Use when creating, editing, reviewing, or finalizing analytical notebooks whose code, assumptions, transformations, diagnostics, or conclusions need to be understandable to another analyst
Use when notebook results depend on parameters, caches, run order, hidden state, seeds, external helper code, or exported tables and the analysis must be rerunnable by another analyst
Use when an experiment analysis, metric study, notebook investigation, or methodology proposal is close to decision-ready and you need an independent check of design, statistics, interpretation, and reproducibility
| name | ds-verification-before-completion |
| description | Use when about to claim that an experiment result, metric validation, or notebook analysis is final, significant, trustworthy, or decision-ready and you need fresh rerun evidence instead of stale outputs |
No analytical conclusion without fresh evidence. If the notebook was not rerun under the stated parameters, or the outputs do not match the written conclusion, the work is not done.
If reruns are brittle, cached, or order-dependent, use ds-notebook-reproducibility first.
Use the shared validation budget from ds-using-superpowers. Intermediate progress claims can use scoped verification. Final, significant, trustworthy, or decision-ready claims require strict verification.
Core principle: Evidence before claims, always; scope the evidence to the claim unless the claim is final or decision-ready.
Violating the letter of this rule is violating the spirit of this rule.
NO FINAL OR DECISION-READY CLAIMS WITHOUT FRESH STRICT VERIFICATION EVIDENCE
If you have not run the verification in this message, you cannot claim the result is trustworthy.
For intermediate progress updates, verify the relevant cells, query, or artifact before making the scoped claim. Do not imply the whole analysis is complete.
Before any final, significant, trustworthy, or decision-ready claim:
Skip any step = not verified.
For intermediate scoped claims:
| Claim | Requires | Not Sufficient |
|---|---|---|
| Experiment is significant | Fresh effect size, interval, and significance output | Old screenshot or p-value alone |
| Metric is valid | Fresh denominator, unit, and robustness checks | “It looked stable earlier” |
| Notebook is reproducible | Clean rerun with explicit parameters | One successful run in dirty state |
| Result is decision-ready | Primary result, limitations, and review completed | Best slice looks good |
| Agent finished task | Artifacts checked independently | Agent summary alone |
| Excuse | Reality |
|---|---|
| "It ran yesterday" | Rerun it now |
| "I am confident" | Confidence is not evidence |
| "The effect is obvious" | Quantify it fresh |
| "Partial rerun is enough" | Partial reruns prove little |
| "The reviewer already checked it" | Verify independently |
Result claim:
✅ [Rerun notebook or script] [See: updated estimate, CI, sample sizes] "Result is X"
❌ "Should still be X"
Requirements claim:
✅ Re-read plan or decision question → verify each required artifact → report actual state
❌ "Notebook ran, so analysis is done"
Delegated work:
✅ Agent reports success → inspect artifacts → rerun critical checks → report actual state
❌ Trust agent report
ALWAYS before:
Use scoped verification before intermediate status claims. Use strict verification before final claims.
Rerun. Read the outputs. Then make the claim.
This skill verifies notebook analysis through reruns, output inspection, and analytical checks, not through unit-test or pytest conventions unless explicitly requested.