ORO Post-Training

Flywheel

Weekly reference-model training on live race trajectories. Every run scored on the fixed eval set v1, and the gate blocks any publish that would regress the deployed model.

Model performance · week over week

success_rate field_matching format_score published

success_rate · latest

—

field_matching · latest

—

since first run

—

Runs

Week	Stages	success	field_match	Gate

Data volume · DPO pairs (balanced)

Published model

run—

success_rate—

eval set—

checkpoint—

gatepublished

Try the model

stretch · coming soon

Bring your own HuggingFace token, pick a shopping problem, and watch the trained agent run it end-to-end — model on your inference endpoint, our search + tool loop, full trajectory and recommendation. Runs on your account, not ours.

Reads run manifests from S3 · AWS-hosted · access-gated. No agent code or secrets shown.