dhresearch/quality-router-v2-utility-maximiser

Utility maximiser (ridge cost, interface features)

This file is a seed-0 refit made on 2026-10-03. The original training run wrote predictions and did not save weights. This upload is the refit.

quality_router_v2 version 8 sets learned_downrouting to false. The served card loads no fitted weights. This file is a learned model from the same study, and the served card leaves it unloaded.

Role

Per-route success heads plus a ridge cost head on the pre-routing numeric block and the interface one-hot. The interface indicator is validator ground truth. Under the failure budget the validation rule stays on strong_repair. The served card does not load this file.

Data

Measured pool https://huggingface.co/datasets/dhresearch/outcome-router-v2-measured-pool: pilot_tasks.jsonl, pilot_outcomes.jsonl, and splits.json.

Fit

fit_heads(..., cost_model='ridge', seed=0) on the same interface rows as the constant-cost fit. This is revalidate_claims._utility_report.

Seed 0. Software at refit time: python 3.14.2, sklearn 1.9.1, numpy 2.5.3, joblib 1.6.0, lightgbm 4.7.0, scipy 1.18.1.

Agreement

Rows: train 602, val 154, test 244. Routes: cheap_model, medium_model, code_specialist, code_specialist_repair, strong_model, strong_repair. Validation cost MAE matches data/real_v2/revalidation_report.json under utility_maximiser.validation_cost_mae, including strong_repair ridge MAE 0.00275026 against the training-mean MAE 0.00356339 (n=154). The validation rule stores no weight: realised failure stayed outside incumbent failure plus 0.01. The original run did not save weights. This file is the seed-0 refit.

Load

joblib.load returns the fit_heads dict. Score with predicted_outcome.predict_routes. A constant head has no model entry.

Class: dict from predicted_outcome.fit_heads.