transport_order_project_transport_strategy_set_ranker-production

Transport order strategy ranking

September 2026 corrected-label Random Forest refresh. Trained from scratch on 68,655 completed managed orders, using the existing 15 model features and 40 eligible candidate strategies. Recent customer/lane-history features are excluded.

Model SHA256: 6f77f5e8e1b9ec6ec870c00b3468a89a39f8ba31e5b8fc8e9e227582a203ef0f.

The API accepts the existing nine order inputs: billable_miles, bearing_deg, pickup_day_of_week, pickup_month, category_id, office_id, customer_id, pickup_location_id and delivery_location_id. Pass them in inputs; optional top_n limits the returned ranking. Responses contain ranked strategy IDs and model scores. Scores are not calibrated probabilities of final selection.

Historical development evaluations

Each window uses a separate model trained before that window, not this full-data artifact evaluated on its own training examples. These historical replays use shared exported inputs and have already been used for development.

The larger window has a first-choice regression. Performance varies by office and lane. test_auc in the configuration comes from the separate September 7 validation fit; it is neither full-artifact holdout accuracy nor top-choice accuracy.

Serving

Keep rf_model.pkl, rf_config.json and strategy_stats.json together. The custom handler uses scikit-learn 1.7.2. Publish the repository revision, then explicitly update the inference endpoint to that exact commit and verify live responses. The source revision and snapshot identity are in rf_config.json.