Anonymous data and reproducibility package for a multisite randomized study of generative-AI scaffolding

Abstract

This anonymized dataset and reproducibility package accompanies a multisite, two-stage randomized study examining the sequencing of learner-generated judgment and generative-AI advice, scaffold withdrawal, and delayed AI-free reasoning. The study involved 3,600 undergraduate business and management students across eight institutions and 45 teaching cohorts. The package contains de-identified scored data, data dictionaries, measurement materials, scoring documentation, AI-system configuration, controlled-error exposure records, randomization and process/fidelity audit files, and executable code for reproducing the primary, secondary, sensitivity, form-specific, missing-data, and process analyses. Direct identifiers, author information, institutional identities, row-level student prompts, model replies, and provider-native request metadata are not included. The package is provided in anonymized form for reproducibility and peer-review purposes.