An analytics skill that's allowed to say "don't deploy"
A reusable agent skill that forces high-stakes data work to define the decision first, gate the data, and treat "no recommendation" as a valid answer.
What it does
High-Stakes Analytics & Decision Lab is a Python-based analytics skill for agents: you give it a decision question and data, and it produces an Evidence Intelligence Report where every claim links back to its source, calculation, and figure. If the evidence actually justifies acting, it adds a separate Decision Intelligence Brief covering alternatives, tail risk, and reversal conditions. It ships with 15 complete public-data cases — 25 intelligence products and 119 accessible figures — as worked precedents.
The interesting bit
The data gate can stop everything. Uploaded data doesn’t go straight into a model: the system preserves the original, checks grain and keys, profiles quality and privacy, and pauses for explicit approval before any imputation, deletion, or outlier treatment. Equally unusual: a valid endpoint can be do_not_deploy, a negative validation, or an evidence request — the skill is built to refuse to recommend when the evidence doesn’t support it.
Key highlights
- Four composable routes (descriptive, diagnostic, predictive, prescriptive), each with required discipline and honest endpoints
- Every figure carries a title, description, linked question, and claim boundary — visuals are part of the evidence contract, not decoration
- The spatial equity case shows the method in action: median imputation vs. zero-filling of missing rent-to-income data changes one of five site selections
- Compact installable bundle (~39 files, ~472 KiB) with a hash manifest, separate from the full research portfolio
- Cases intentionally end differently — some bounded decisions, some pilots, some stops
Caveats
- The “skill” framing means it’s designed for agent runtimes (with no-install and direct-repository options documented); it’s not a standalone library in the usual sense
- The README is dense — the architecture docs are doing a lot of load-bearing work
Verdict
Worth a look if you build agent workflows where analyses affect real people and budgets, or if you want a template for reproducible, decision-first reporting. Skip it if you just want a quick dataframe-to-chart pipeline.
Frequently asked
- What is limingrui679-design/high-stakes-analytics-decision-lab?
- A reusable agent skill that forces high-stakes data work to define the decision first, gate the data, and treat "no recommendation" as a valid answer.
- Is high-stakes-analytics-decision-lab open source?
- Yes — limingrui679-design/high-stakes-analytics-decision-lab is open source, released under the MIT license.
- What language is high-stakes-analytics-decision-lab written in?
- limingrui679-design/high-stakes-analytics-decision-lab is primarily written in Python.
- How popular is high-stakes-analytics-decision-lab?
- limingrui679-design/high-stakes-analytics-decision-lab has 1k stars on GitHub.
- Where can I find high-stakes-analytics-decision-lab?
- limingrui679-design/high-stakes-analytics-decision-lab is on GitHub at https://github.com/limingrui679-design/high-stakes-analytics-decision-lab.