lemma

Establish a dumb baseline and an honest validation harness before any real model, so every later number means something.

Rigor for causal questions and A/B tests (the effect of acting on X): confounding, post-treatment bias, valid control groups.

Rigor for descriptive and diagnostic analytics (what happened and why): denominators, grain, and confounded slices, not model leakage.

EDA kickoff for a fresh dataset: fixed opening scaffold (goal, imports, load, sanity), then chapters derived from the data; scan leakage, land a baseline.

Rigor for statistical inference (is the difference real): hypothesis tests, power, multiple comparisons, effect size over p-value.

Audit a dataset or pipeline for the five leakages that inflate a metric: target, preprocessing, temporal, group, and sampling.

Final modeling once the baseline and feature set are locked: tune against validation, audit overfitting, touch the test set once, justify the complexity.

Review a notebook or analysis for data-science anti-patterns before it's trusted or shared.

Rigor for clustering, dimensionality reduction, and anomaly detection: validity is stability under resampling, not a held-out score.

Assemble a trustworthy working dataset from messy or multiple sources: grain, keys, joins with match rates, extraction checks, lineage.

Establish an honest score to beat before complex modeling, including the validation design, metric, no-information rule, and simplest credible model.

Estimate the effect of an intervention for experiments and defensible quasi-experimental or observational designs; do not substitute prediction for identification.

Use for complex descriptive decompositions such as cohorts, funnels, segment comparisons, and what-changed investigations. Skip bounded lookups, joins, rankings, counts, averages, and aggregates.

Explore a fresh dataset when the analytical direction is open; use for orientation, pattern discovery, and deciding what analysis is worth pursuing.

Quantify whether a difference or association is distinguishable from sampling noise using effect estimates, uncertainty intervals, tests, or power analysis.

Audit suspicious model performance or a pipeline for target, preprocessing, temporal, group, sampling, or duplicate contamination.

Select and evaluate a production-worthy model after an honest baseline and validation design exist; use for tuning, calibration, thresholding, and final evaluation.

Review a notebook or analysis for correctness, reproducibility, leakage, weak validation, unsupported claims, and misleading communication.

Discover or evaluate structure without labels, including clustering, anomaly detection, embeddings, dimensionality reduction, and topic models.

Reconcile sources into a defensible analytical dataset when grain, keys, definitions, units, authority, extraction, joins, or provenance are uncertain.