Why This Policy Exists

Reproducibility in AI/ML research depends on more than the underlying data — code, trained model parameters, and the exact training procedure all materially affect whether a result can be independently verified. This policy sets expectations beyond a general data-sharing statement.

Data Availability Statement

A Data Availability Statement is required in every manuscript, accurately describing whether and how the underlying data can be accessed. Acceptable statements include: openly available in a repository with a DOI; available upon reasonable request with access conditions stated; subject to institutional, legal, or patient-privacy restrictions, with the restriction explained; or secondary/public datasets, identified and cited.

Code Availability

Authors are strongly encouraged to share the code used to train and evaluate their model — including preprocessing scripts, model architecture definitions, and evaluation scripts — in a trusted public repository (for example, GitHub with an archived release, or Zenodo). Where code cannot be shared, authors state the reason and describe the method in sufficient detail for independent re-implementation.

Trained Models and Weights

Where legally and ethically permissible, authors are encouraged to share trained model weights or parameters, or to specify the exact conditions under which they may be obtained. Where a model cannot be shared due to licensing, commercial, or patient-privacy constraints, this is stated explicitly.

Human Participant and Patient Data

Identifiable patient data must not be deposited in an open repository. De-identified data may be shared only when consistent with consent, ethics approval, data-protection requirements, and applicable law, consistent with the Informed Consent Policy.

Reproducibility Expectations

Authors provide sufficient methodological, data, and code detail to allow a qualified researcher to understand and assess the analysis. Sharing data, code, or model weights is strongly encouraged but does not override privacy, confidentiality, security, intellectual-property, or contractual restrictions.