Results

Surfaces

• Experiment runs — browse reports/runs/, open manifests/metrics, compare up to four runs, promote to active model
• Charts — model compare, selected metric, feature importance (Chart.js)
• Model comparison — table of multi-model runs
• Metrics — filterable hold-out metrics (AUROC, PR-AUC, Brier, ECE, …)
• Fairness — age-band / group metrics table (run from Results; not ZIP-only)
• Operating points — precision/recall at several probability thresholds (research only)
• Light HPO — optional grid trials from Train, presented as a structured best-trial card and trial table; unavailable metrics render as n/a, not raw JSON
• Feature importance — top drivers from the fitted model
• Leakage audit — findings from the audit job (bind run_id)
• External validation — score a held-out / site CSV against the run model
• Trust checklist — per-run trust pack completeness
• Report files — JSON/CSV artifacts under reports/
• Figures — calibration and SHAP PNGs
• Paper curves — interactive ROC / PR / calibration live on Analytics after retrain

Actions

• Generate SHAP — background job for summary explanations
• Run fairness — subgroup metrics on hold-out
• Threshold table — compute operating points from the active artifact
• Download ZIP — pack metrics, audits, fairness/HPO (when present), figures
• Empty state links back to Train when no model exists

Next: Predict · Analytics