Data & unit checks
Verify sample counts, units, labels, exclusions, missing values, duplicates and cohort flow before interpreting any output.
WORKFLOW STAGE 03
Results & validation.
Convert raw outputs into defensible evidence through correct metrics, statistical inference, robustness checks, transparent comparisons and disciplined error analysis.
EvaluationStatisticsRobustnessError Analysis
Validate the claim, not only the score.
A strong result shows what improved, how reliably it improved, where it fails and whether the comparison is fair.
Baseline parity
Evaluate all methods on identical splits, preprocessing, resources and metrics. Report both strong and conventional baselines.
Metric suitability
Match measures to the task, class balance, decision cost and research objective; avoid conclusions based on one convenient metric.
Statistical evidence
Report confidence intervals, effect sizes, assumptions and suitable tests. Correct for repeated or multiple comparisons when needed.
Robustness & sensitivity
Test alternative seeds, thresholds, hyperparameters, subsets, perturbations and external data; use ablation to isolate each component.
Error analysis
Inspect failure cases, subgroup performance, residuals, confusion patterns and outliers, then connect limitations to evidence.
Validation by research type.
Use measures that reflect the actual scientific question and data structure.
| Classification | Confusion matrix, precision, recall, F1, ROC-AUC or PR-AUC, calibration, class-wise scores and confidence intervals. |
|---|---|
| Image segmentation | Dice, IoU, sensitivity, specificity, Hausdorff distance, boundary errors, visual cases and external-dataset testing. |
| Regression | MAE, RMSE, R², residual diagnostics, prediction intervals, heteroscedasticity checks and practical error limits. |
| Time series | Walk-forward validation, horizon-wise error, MASE or sMAPE, temporal drift and naive or statistical benchmarks. |
| Simulation / CFD | Convergence history, grid independence, conservation checks, uncertainty quantification and benchmark or experimental agreement. |
| Survey / SEM | Reliability, convergent and discriminant validity, fit indices, path estimates, mediation/moderation and sensitivity analysis. |
Avoid: reporting only the best run, treating statistical significance as practical importance, omitting negative results, comparing unequal settings, or claiming generalization without independent validation.
Need stronger validation for your results?
Share your outputs, baselines and target claims for a focused technical review.