dhresearch/quality-router-v2-domain-code-policy

Domain code policy (quality router v2 phase)

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

Route-conditioned success model. Code interfaces fall back to code_specialist_repair. Candidates are tool_route, cheap_model, medium_model, and strong_model. This is the default domain policy in scripts/run_real_v2.sh. The served flow does not call this object. Economy mode on the served card names a learned downrouter and is off.

Data

Training rows are the prompt-only export of the 1,000-task pilot, split 600/150/250. Dataset: https://huggingface.co/datasets/dhresearch/outcome-router-v2-main. Local directory used for this refit: data/real_v2/main_router_prompt_only.

Fit

python3 -m outcome_router_data.scripts.train_router --dataset data/real_v2/main_router_prompt_only --router domain_code_policy --routes configs/routes.pilot_v2_real.yaml --candidate-routes tool_route cheap_model medium_model strong_model --split test. The saved object is the fitted RouteConditionedSuccessRouter.

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

Frozen file data/real_v2/preds_v2_domain.jsonl: predicted_route matches on 250 of 250 shared test rows. Full rows, including rounded probabilities and labels, match on 250 of 250. The original run wrote that prediction file and did not save weights. This upload is a seed-0 refit that reproduces those test rows. It is not a copy of an original weight file.

The current quality_router --economy command builds this same router. Its refit has the same sha256. See dhresearch/quality-router-v2-economy-downrouter for the frozen economy file, which this object matches on 156 of 250 test routes.

Load

Unpickle with joblib while outcome_router_data is importable. predict_rows expects route_outcome_matrix rows.

Class: RouteConditionedSuccessRouter.