bluegreencity-ki

ML Analysis Reports

This folder contains finalized analytical reports for the ML risk assessment branch.

Current Reports

• treelife_overlap_report_2026-10-09.md - External validation of the 2026_10 retraining against real LHH Treelife interventions (2024 holdout and operational 2026 scores): consistent 1.4–2.0× enrichment of treated trees in the model’s top-risk ranking; a useful soft monitoring signal, not an acceptance gate
• operational_panel_grouped_cv_report_2026-09-16.md - Latest published grouped 5-fold robustness check for the operational panel model with objectid-based fold grouping (September 2026 retraining round, v6 panel, 35 features); the October 2026 round’s grouped-CV results are in the model directory, successor report pending. Successor of the June 2026 report, which is now archived
• stable_ablation_report_2026-04-03.md - Comparison of the stable AutoGluon reference run and the documented feature-ablation variants
• high_risk_feature_combinations_2026-04-07.md - Analysis of the feature combinations and local drivers behind the highest ML risk in the current production result
• baseline_vs_minimal_dashboard_comparison_2026-04-07.md - Comparison of the baseline dashboard/run and the new minimal-run dashboard, including model validation, feature shifts, and 2025 output differences
• operational_panel_result_report_2026-04-08.md - Compact comparison of the operational panel model against baseline and minimal, including 2024->2025 backtest validation and critical assessment

Scope

Use this folder for compact, date-stamped result reports that interpret a concrete experiment or production result.

Use ../archive/README.md for older planning notes, exploratory drafts, or superseded implementation summaries.