Develop and compare predictive models with realistic baselines, deployment-matched validation, bounded tuning, and reproducible experiments while keeping the final test untouched.
