inst/governance/GOVERNANCE.md

In gp3ml: Governance-First Predictive Modelling for 'Gazepoint' Research

gp3ml Model Governance

Purpose

gp3ml is intended to provide leakage-resistant, group-aware, auditable predictive-validation infrastructure for explicitly labelled, non-sensitive outcomes in Gazepoint research workflows.

The package is not intended to serve as a generic machine-learning training wrapper.

Required generalization target

Every predictive evaluation must declare one of the following:

• new trials from known participants;
• new participants;
• new stimuli; or
• new participants and new stimuli.

The data split and resampling structure must be compatible with the declared target.

Core safeguards

• Participant leakage must be prevented by default.
• Stimulus leakage must be prevented when unseen-stimulus generalization is claimed.
• Random row-level splitting is not the default.
• Preprocessing must be estimated inside resampling folds.
• Outcome-derived feature selection is prohibited.
• Identifiers must not be used as predictors.
• Evaluation must not rely on accuracy alone.
• Calibration and uncertainty must be reported when applicable.
• Synthetic data used in examples and tests must be deterministic.

Initial model scope

The initial package will not provide broad model-training wrappers. Transparent baseline models may be considered only after leakage auditing, grouped resampling, evaluation, and reporting safeguards are complete and validated.

gp3ml documentation built on Aug. 23, 2026, 5:11 p.m.