mirror of
https://github.com/magnus919/agent-skills.git
synced 2026-09-21 00:26:23 +03:00
Magnus HedemarkandGitHub
4ea078ec7e
* feat(skill): add interpretability workflow * fix(data-scientist): make the causal route clickable * docs(data-scientist): list all interpretation resources
2.1 KiB
2.1 KiB
Interpretability workflow
Interpretability is evidence about model behavior under a method and reference distribution. It is not automatically a causal explanation or a user-facing justification.
- State the decision, audience, stakes, and explanation target: global model behavior, local prediction, error diagnosis, fairness investigation, or a contrastive question. Define whether the audience needs a diagnostic view or an actionable explanation.
- Choose scope and background deliberately. Record the model, data version, reference population, feature preprocessing, missingness, and perturbation or intervention semantics. For local explanations, state the neighborhood and baseline; for global explanations, state the population and aggregation.
- Check correlated features, proxies, extrapolation, and distribution shift. A feature attribution may be shared among correlated variables or reflect a model shortcut. A perturbation may create impossible records and should be marked invalid rather than interpreted.
- Validate the explanation: rerun with seeds, nearby backgrounds, and relevant perturbations; use label or feature randomization sanity checks where applicable; compare at least one materially different method or model. Record disagreement instead of averaging it into false certainty.
- Inspect slices, especially protected groups and high-impact cases. An aggregate explanation or fairness result can hide a subgroup failure. Report sample sizes, uncertainty, missingness, and the limits of slice comparisons.
- Write the conclusion in predictive language unless a causal design identifies an intervention effect. Never turn “the model relied on X” into “X caused Y.” Route causal claims to the causal-inference framework.
Explanation outputs belong in a versioned report with method, target, audience, background, stability results, method disagreement, slice results, and known failure modes. Do not expose sensitive feature values or internal reasoning just to make an explanation look complete.