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

gp3ml-package

R Documentation

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

Description

gp3ml provides governance-first infrastructure for leakage-resistant predictive modelling and validation using Gazepoint-derived research data.

Details

Core capabilities include explicit task and variable-role declarations, feature-provenance manifests, leakage auditing, group-aware holdout splitting and repeated resampling, fold-local preprocessing, governed model engines, performance and calibration assessment, target-aligned uncertainty, external-validation reports, prediction-to-decision governance, target-aware conformal prediction, dataset-shift auditing, locked analysis plans, portable model artifacts, robustness diagnostics, environment provenance, research-object export, model cards, and reproducibility evidence.

Repository-aware evaluation and tuning

Materialized gazepoint_group_folds can be evaluated without rebuilding or replacing the fold object. Preprocessing and model fitting occur only within each analysis partition; predictions are produced only for the matching assessment partition. Explicit candidate grids retain failed candidates and require a declared metric, direction, and human rationale before selection.

Nested resampling and uncertainty

Nested grouped resampling isolates inner tuning inside each outer analysis partition. Target-aligned uncertainty distinguishes observation, participant-cluster, stimulus-cluster, simultaneous participant/stimulus, fold-distribution, and repeat-distribution summaries. An uncertainty object may not be described as uncertainty for an undeclared unit.

External validation

External validation requires an explicit independent-dataset declaration. Reports include predictor availability, schema differences, prevalence shift, calibration drift, participant/stimulus coverage, and transportability limitations. Internal holdouts remain explicitly labelled as not externally validated.

The package is intended only for explicitly observed, non-sensitive outcomes and declared scientific purposes. It does not support person identification, biometric authentication, health or protected-attribute inference, or direct or indirect inference of emotion, stress, personality, deception, cognition, comprehension, intent, or other mental states.