Stress-Test Analytical Conclusions for Decision Robustness

Required inputs

Have these details available before following the usage instructions.

• The analytical conclusion or recommendation being tested and the decision it may support.
• Data sources, extraction period, population, cohort definitions, inclusion and exclusion rules, metric definitions, and known transformations.
• Model, query, dashboard, or analysis method used, including assumptions, filters, joins, weighting, imputation, segmentation, and aggregation logic.
• Observed effect sizes, confidence or uncertainty measures if available, historical baseline, comparator, and practical decision threshold.
• Known data-quality issues, missingness, outliers, identity-resolution concerns, timing effects, or measurement changes.
• Alternative plausible definitions, cohorts, assumptions, model choices, and stakeholder concerns to test.
• Decision owner, risk tolerance, consequences of false positives and false negatives, and deadline or reversibility constraints.

How to use this Skill

When to use:
- Before relying on an analytical conclusion for an executive, product, finance, policy, or operational decision.
- When a metric movement, forecast, model result, dashboard insight, or segment finding appears consequential but may depend on assumptions.
- When stakeholders disagree about definitions, cohorts, missing data, exclusions, model choices, or interpretation.
- When preparing a decision brief that must disclose robustness, uncertainty, and conditions for use.

When not to use:
- Do not use to validate raw data lineage incidents; use a data-lineage investigation first when source divergence is suspected.
- Do not use as a substitute for executing statistical tests, rerunning pipelines, or obtaining specialist review for high-stakes statistical claims.
- Do not use when the conclusion, decision context, source data, definitions, and assumptions are not available.
- Do not use to force a binary answer when the evidence supports only limited or conditional use.

Instructions:
1. Open AMO-P-000319, supply its required inputs, and run the complete source Prompt in ChatGPT. Review its evidence register and limitations before using the result.
2. Restate the conclusion, intended decision use, supplied uncertainty and practical decision threshold, and consequences of being wrong. Missing confidence requirements or thresholds remain unresolved; do not invent them.
3. Build an evidence ledger that separates supplied facts, assumptions, inference, missing information, and uncertainty.
4. Map the conclusion’s dependency chain: source data, eligibility rules, metric definitions, transformations, model choices, segment cuts, missing-data treatment, and interpretation steps.
5. Identify plausible sensitivity dimensions, such as cohort window, denominator choice, outlier handling, attribution rules, lag assumptions, seasonality, identity joins, missingness, model specification, and segment aggregation.
6. For each sensitivity dimension, state the expected direction of risk, the evidence needed to test it, available results if supplied, and whether it could overturn, weaken, strengthen, or limit the conclusion.
7. Distinguish robustness findings from recommendations. Provide decision-use language such as use, use with limits, recompute, investigate, hold, or not enough evidence.
8. Require human review for consequential, financial, legal, regulatory, production, safety, or publication decisions based on the conclusion.

Expected output:
A decision-robustness brief containing the conclusion under review, evidence and assumption ledger, sensitivity matrix, possible overturning conditions, unresolved data or method risks, recommended follow-up checks, and an evidence-qualified decision-use disposition.

Constraints and boundaries:
- Do not invent statistical results, rerun analyses, or imply access to data not supplied by the user.
- Do not treat robustness as proof of causality unless the design and evidence support causal interpretation.
- Do not hide uncertainty or collapse conflicting evidence into false certainty.
- Do not provide financial, legal, medical, regulatory, or safety approval; provide analytical decision support for qualified review.
- Keep sensitivity recommendations feasible and aligned with the decision’s urgency and reversibility.
- Keep arithmetic checks, specification sensitivity, sampling uncertainty, causality, external validity and data-quality exposure separate. An unchanged sign does not establish robustness when decision magnitude or affected population changes. Proposed reruns cannot establish a robustness result.

Based on an Amo.ng Prompt

Analytical Conclusion Sensitivity Review

Open the linked Prompt to read the instructions this Skill is based on.

Completion criteria

Complete when:
- The reviewed conclusion is tied to a specific decision, threshold, and consequence profile.
- Each sensitivity risk is linked to a data, definition, assumption, model, or interpretation dependency.
- The output identifies which changes could materially alter the conclusion and which are unlikely to matter, with evidence limits stated.
- Missing information and unrun checks are clearly separated from observed results.
- The final disposition is conditional on the evidence and does not overclaim reliability.