Changelog • gp3ml

gp3ml 0.2.0

Added fold-local evaluation across materialized gazepoint_group_folds, with retained predictions, metrics, exclusions, leakage audits, warnings, and failures.
Added explicit deterministic tuning grids, candidate comparison, and human-reviewed selection requiring a predeclared metric and direction.
Added nested grouped resampling with inner tuning isolated inside every outer analysis partition.

Added observation, participant-cluster, stimulus-cluster, and participant-and-stimulus bootstrap uncertainty with the resampling unit recorded explicitly.
Added fold- and repeat-distribution summaries without relabelling row-level results as cluster-level estimates.

Added independent-dataset declarations, schema checks, predictor availability, prevalence shift, calibration drift, participant/stimulus coverage, and transportability limitations.
Added explicit not_externally_validated status when no independent dataset is available.
Added release-ready model cards recording selection procedures, uncertainty units, transportability status, limitations, and human oversight.

Added nine deterministic synthetic vignettes covering recording quality, assigned conditions, observed outcomes, participant and stimulus generalization, simultaneous participant/stimulus generalization, contaminated manifests, nested resampling, and external validation.
Added focused tests, an end-to-end smoke test, a full validation driver, and a read-only 0.2.0 release audit.

Established gp3ml as a governance-first package for leakage-resistant predictive modelling and validation using Gazepoint-derived research data.
Restricted supported tasks to explicitly observed, non-sensitive outcomes and clearly declared scientific purposes.
Added explicit prohibited-use documentation covering identification, authentication, diagnostic, protected-attribute, emotion, stress, personality, deception, cognition, comprehension, intent, and other mental-state inference.

Added task declaration, use-case assertion, variable-role validation, and machine-readable prohibited-use helpers.
Added feature-provenance manifests covering predictor origins, transformations, availability stages, roles, and preprocessing scope.
Added structured leakage audits for row, participant, participant-trial, stimulus, identifier, target-derived, and post-outcome risks.

Added deterministic group-aware holdout splitting and repeated grouped resampling for new trials among known participants, new participants, new stimuli, and simultaneous new-participant and new-stimulus generalization.
Materialized analysis, assessment, and explicitly excluded partitions with complete source-row accounting and embedded leakage audits.
Added fold-balance, coverage, exclusion, and outcome-representation diagnostics with structured pass, review, and fail findings.

Added fold-local preprocessing objects with separate fitting and baking interfaces.
Added governed model engines, explicit black-box integration, classification and regression metrics, calibration assessment, and explicitly labelled bootstrap metric uncertainty.
Added external-validation evaluation and reporting without treating an internal holdout as external validation.
Added model cards and reproducibility reports that record task purpose, governance decisions, model settings, performance, calibration, uncertainty, limitations, and reproducibility information.
Model selection remains explicit and reviewable; the package does not perform autonomous black-box winner selection.

Added machine-readable CSV, Markdown, and JSON reporting interfaces where supported by the relevant object.
Added deterministic synthetic tests across supported generalization targets, governance failures, leakage cases, model engines, metrics, calibration, reporting, and serialization.
Added complete pkgdown reference organization for the first formal release.