The development API is deliberately frozen during this hardening phase. The purpose is to deepen integration and maintenance guarantees rather than add new analytical surface area.
A machine-readable manifest records all exported function names and formal argument names. Tests compare the installed namespace against this manifest so accidental public additions, removals, or signature changes become explicit failures rather than silent drift.
The relevant governance interfaces include
validate_gp3bayes_object(),
capture_gp3bayes_schema(),
validate_gp3bayes_schema(),
create_analysis_manifest(), and
compare_analysis_manifests().
Small adapters are useful because they let downstream reports use
stable data frames rather than inspect internal object fields. Examples
include backend_environment_table(),
loo_influence_atlas_table(),
prediction_profile_table(),
prediction_surface_table(),
prediction_draws_long(), and
prior_posterior_draws_long().
Fit-dependent extraction helpers such as
extract_expected_predictions(),
extract_posterior_predictions(),
extract_linear_predictions(),
extract_log_likelihood(), and
extract_sampler_diagnostics() reject malformed inputs
rather than guessing.
Prediction-comparison helpers retain explicit bounds. In particular,
prediction_pairwise_contrasts() and
prediction_rank_probabilities() require the analyst to opt
into larger comparison sets rather than expanding combinatorially
without review.
The diagnostic layer separates descriptive posterior evidence from
automatic decisions. binary_group_calibration(),
posterior_predictive_summary_table(),
predictive_coverage_table(),
duration_pit_table(), and
loo_group_influence_table() return evidence for review;
none automatically certifies adequacy or excludes
observations/groups.
Writers such as write_model_card(),
write_publication_registry(),
write_diagnostic_dashboard_report(),
write_analysis_bundle_report(), and
save_publication_registry_figures() remain explicit-output
operations. The package does not use the current working directory as an
implicit reporting destination.
Not every exported wrapper has a runnable Rd example. Many functions require a fitted Bayesian backend object, and duplicating expensive fits across hundreds of help topics would make checks slower without improving the underlying API. The package therefore combines short deterministic Rd examples for lightweight functions with articles, unit tests, integration tests, and complete reference documentation for fit-dependent workflows.