gp3bayes 0.5.0

Advanced dynamic and measurement-aware Bayesian pupillometry

• Adds an additive advanced pupil-model specification layer while preserving the frozen 0.4 public API.
• Adds Gaussian and Student-t observation models with constant, condition-dependent, time-dependent, and condition-by-time residual scale specifications.
• Adds bounded governed residual temporal dependence through AR(1), AR(2), ARMA(1,1), and explicit ARMA specifications constrained to p <= 3 and q <= 2.
• Adds governed Gaussian-process trajectories with Matérn 3/2, Matérn 5/2, and exponentiated-quadratic kernels; approximate Hilbert-space GP bases are the scalable default and exact GP requests are subject to a complexity audit.
• Adds explicit known measurement-uncertainty declarations for pupil responses and covariates and MAR-oriented joint missing-value models without automatic interpolation.
• Adds joint binocular pupil preparation, multivariate modelling, residual eye correlation, eye-difference estimands, and agreement summaries without requiring upstream eye averaging.
• Adds governed PSIS-LOO, exact K-fold, predictive model weights, and explicit leave-future-out validation plans. Model comparison never chooses a substantive winner automatically.
• Adds an experimental nonlinear pupil response-shape model with interpretable baseline, amplitude, onset, rise, duration, and decay parameters.
• Adds posterior temporal derivative, dynamic condition-contrast, and prespecified threshold-duration estimands without automatic onset, changepoint, or favorable-window detection.
• Adds computational-budget, design-support/identifiability, temporal-dependence, measurement, missingness, predictive-calibration, and advanced-fit diagnostics.
• Adds posterior residual-scale, GP hyperparameter, residual-spectrum, model-card, sensitivity-suite, and publication-oriented tables and graphics.
• Adds deterministic advanced and binocular simulation utilities with truth stored separately from analysis data.
• Adds backend-free examples, plotting galleries, focused tests, and eleven advanced pupillometry articles.

Governance boundaries

• gp3bayes 0.5 does not interpolate or otherwise preprocess missing pupil samples automatically.
• MAR-oriented missing-data modelling is assumption-conditional and does not establish that MAR is true.
• Student-t robustness does not mark individual observations as invalid outliers.
• Predictive comparison, stacking weights, temporal derivatives, threshold durations, and nonlinear response parameters do not establish model adequacy, causality, or cognitive/physiological states.
• MNAR models, overlapping-event deconvolution, automatic changepoint detection, multimodal psychophysiological latent-state inference, automated model search, and automatic cognitive-state inference remain outside the 0.5 scope.

gp3bayes 0.4.0.9000

Bayesian dynamic pupillometry foundation

• Adds a first-class, vendor-neutral pupil time-course contract with verified Gazepoint schema inspection and explicit unit handling.
• Adds deterministic pupil simulation, governed preparation, readiness and measurement-context audits without silent blink interpolation, smoothing, PFE correction, luminance correction, or exclusion.
• Adds a restricted Gaussian hierarchical pupil time-course family with smooth or linear trajectories, optional condition-specific trajectories, participant/item hierarchy, declared covariates, scale-aware priors, and an optional guarded AR(1) structure.
• Adds backend-portable full-MCMC fitting through brms with rstan or cmdstanr, plus prior-predictive planning/execution under the same closed specification.
• Adds posterior pupil trajectories, declared-window means, AUC, peak response, peak latency, condition contrasts, threshold probabilities, PPCs, temporal diagnostics, target-specific grouped/future validation, and declared sensitivity scenarios.
• Adds publication tables, ggplot2 graphics, nine pupillometry articles, focused failure-contract tests, a frozen 0.4.0.9000 API manifest, and a dedicated development audit.
• Pupil responses are not automatically interpreted as cognitive load, attention, arousal, stress, emotion, surprise, or effort; causal, adequacy, exclusion, and model-selection claims remain outside automated package decisions.

gp3bayes 0.3.0.9000

Public API and integration hardening

• Froze the 324-function development API in a machine-readable manifest and added tests for exported names and formal argument stability.
• Added source-level documentation coverage tests for all public exports.
• Added direct smoke tests for lightweight prediction, backend, LOO, prior-posterior, and Phase-4 table adapters.
• Added explicit malformed-input and combinatoric-boundary tests for fit-dependent extraction and prediction-comparison helpers.
• Added a complete public API map and a quality/failure-contract article.

This hardening phase adds no public functions and does not broaden the approved model-family scope.

Hierarchical and posterior-predictive atlases

• Added raw group-effect draw extraction, posterior rank probabilities, and baseline random-intercept latent variance partitions.
• Added governed numeric prediction profiles, finite-difference predictive gradients, two-dimensional prediction surfaces, and contrast profiles.
• Added posterior-predictive distribution and quantile atlases, posterior uncertainty in prediction scores, and binary calibration uncertainty.
• Added group-aggregated PSIS-LOO influence summaries and graphics.
• Added four articles covering the new hierarchical and predictive layer.

All additions remain descriptive under the fitted model. They do not add causal derivatives, automatic ranking, automatic calibration certification, automatic adequacy decisions, or automatic group exclusion.

Evidence atlases, recovery graphics, and publication registries

• Added publication-oriented recovery, prior-sensitivity, estimand- sensitivity, group-deletion, random-slope, power-scale, and SBC adapters.
• Added a declared-prior versus posterior bridge with marginal shift, contraction, overlap, and empirical distance summaries.
• Added pointwise PSIS-LOO influence atlases without automatic exclusion.
• Added publication registries, evidence inventories, and non-interactive diagnostic dashboards with explicit file-output semantics.
• Added seven evidence/publication articles including an end-to-end showcase.

No addition performs automatic model selection, automatic exclusion, automatic adequacy or robustness certification, or causal interpretation.

Advanced predictive diagnostics and evidence graphics

• Added ROC, precision-recall, confusion, calibration-error, grouped calibration, predictive Q-Q, duration-tail, interval-width, posterior ranking, and posterior predictive discrepancy summaries.
• Added ggplot adapters for sensitivity suites, model-evidence inventories, backend parity/environment checks, analysis-manifest comparisons, schema comparisons, design-support audits, and missingness audits.
• Added structured model cards and reporting-evidence inventories with explicit Markdown output.
• Added four advanced post-fit articles and additional test coverage.

These additions are presentation and diagnostic layers. They do not add automatic model selection, automatic adequacy certification, automatic exclusion, or causal interpretation.

Post-fit exploration, prediction, and publication layer

• Added standardized posterior-draw, sampler-diagnostic, log-likelihood, expected-prediction, posterior-predictive, and linear-predictor extraction.
• Added governed prediction grids and explicit prediction-support auditing.
• Added binary calibration, threshold metrics, predictive scores, duration quantile calibration, PIT summaries, predictive coverage, residual review, grouped posterior predictive checks, and descriptive uncertainty decomposition.
• Added group-effect, variance-component, LOO diagnostic, LOO comparison, and predictive-weight tables.
• Added a publication-oriented ggplot/bayesplot layer covering posterior intervals, densities, MCMC diagnostics, calibration, prediction intervals, hierarchical effects, uncertainty, and LOO influence/comparison.
• Added explicit figure sets and structured analysis bundles; no output is written without an explicit destination.
• Added five articles documenting posterior exploration, prediction and scoring, hierarchical uncertainty, LOO comparison, and publication bundles.

All additions remain within the approved hierarchical Bernoulli-logit and positive uncensored lognormal-duration model families. They do not add automatic model selection, automatic exclusions, adequacy claims, or causal interpretation.

gp3bayes 0.2.0

0.2.0 stabilization program

• Adds a stable family-neutral workflow API while retaining the existing binary- and duration-specific interfaces.

• Adds analysis manifests, data/specification fingerprints, explicit manifest freezing/comparison, and reproducibility reports.

• Adds pre-fit missingness, fixed-effect design, random-effect support, and combined design-support audits without automatic data/model changes.

• Adds declarative unified sensitivity suites and evidence inventories without aggregate robustness, adequacy, exclusion, or selection claims.

• Adds backend environment validation, MCSE-aware rstan/cmdstanr posterior parity auditing, and serialized gp3bayes object-schema contracts.

• Forward-ports CRAN 0.1.1 compliance safeguards: two-core automatic defaults, explicit report paths, temporary vignette outputs, and safe seed handling without direct global-environment modification.

• Adds five focused stabilization articles plus an integrated synthetic 0.2.0 release case study, tests, and release smoke/audit scripts.

• Aligned DESCRIPTION, README, citation metadata, package-level help, backend-installation guidance, CRAN comments, and pkgdown deployment metadata with the complete 0.2.0 API and dual rstan/cmdstanr backend support.

Specification closure

• Adds strict readiness checks for overall condition imbalance, binary group outcome variation, identifier-like numeric predictors, fixed-effect rank, duration extremes, declared duration ranges, censoring signals, and optional separation screening.

• Adds reusable transformation recipes with forward replay, inversion, and exact replay validation for retained rows.

• Adds first-class design-standardised binary probability contrasts and duration median, ratio, and predictive-quantile estimands.

• Adds governed structural, group-deletion, contrast-coding, predictor-scaling, and duration-unit sensitivity workflows without automatic model selection or exclusion.

• Adds detailed family-specific posterior predictive checks and plotting helpers.

• Adds a governed exact K-fold adapter through brms::kfold() as an optional predictive-validation fallback/complement to PSIS-LOO.

• Adds an auditable specification-traceability matrix, examples, smoke tests, and three integrated articles.

• Added optional power-scaling sensitivity integration through priorsense.

• Added conservative PSIS-LOO diagnostics, influence inspection, model comparison, and stacking or pseudo-BMA weights through loo.

• Added fixed-effects separation screening through detectseparation.

• Added simulation-based calibration plans and plots through SBC.

• Added restricted full-MCMC backend selection between rstan and cmdstanr.

• Added coefficient-specific interaction-prior defaults for binary and duration contracts.

• Added dedicated binary and duration pathology generators, evaluations, plots, tests, examples, smoke tests, and three integrated articles.

gp3bayes 0.1.1

• Published gp3bayes 0.1.1 on CRAN after addressing CRAN review feedback.
• Added Zenodo DOI documentation and current release-status wording.
• Added explicit copyright-holder metadata for the initial CRAN submission.

gp3bayes 0.1.0

• Created the independent gp3bayes package scaffold.
• Defined the initial scope as contract-first Bayesian workflows for hierarchical behavioural data.
• Restricted initial development to hierarchical Bernoulli-logit and hierarchical lognormal-duration model families.
• Added package-level documentation and explicit interpretation boundaries.
• Added the initial deterministic scope test using testthat edition 3.
• Added the MIT licence.
• Added standard GitHub Actions workflows for cross-platform R CMD check and pkgdown deployment.
• Added canonical repository, issue-tracker, and pkgdown website metadata.
• Added create_model_contract() for the two approved initial model families with neutral column mappings and explicit methodological specifications.
• Added a concise gp3bayes_model_contract print method and deterministic validation tests.
• Added audit_model_readiness() for backend-independent assessment of outcome validity, declared columns, missingness, repeated measurements, item and trial structure, predictors, interactions, time terms, and requested participant-level random slopes.
• Added structured gp3bayes_readiness_audit results with explicit pass, warning, and failure statuses and a concise print method.
• Added build_model_formula() for deterministic, backend-independent construction of approved fixed-effects, interaction, participant, item, time, and optional participant-level random-slope structures.
• Added create_prior_specification() and validate_prior_specification() for explicit binary-logit and lognormal-duration prior records without creating executable backend objects.
• Added create_model_specification() to combine a model contract, successful readiness audit, approved formula, and validated priors into one inspectable backend-independent specification.
• Added concise print methods and deterministic validation tests for formulas, priors, compatibility checks, and complete specifications.
• Added simulate_hierarchical_binary_data() for deterministic hierarchical Bernoulli-logit simulation with participant effects, optional crossed item effects, optional participant condition slopes, controlled imbalance, and a stored true-parameter record.
• Added prepare_hierarchical_binary_data() for explicit binary-outcome mapping, condition coding, recorded predictor scaling, missing-data decisions, readiness auditing, and fixed-effects matrix construction.
• Added specify_binary_model() to combine prepared data with the approved binary contract, restricted hierarchical formula, and validated backend-independent prior specification.
• Added check_binary_prior_predictive() for deterministic simulation of family-specific prior predictions and structured plausibility checks without fitting a model or requiring a Bayesian backend.
• Added concise print methods, generated documentation, and 89 focused tests for the backend-independent binary workflow foundation.
• Added repository and installed-package citation metadata through CITATION.cff and inst/CITATION.
• Refined the package description to match the currently implemented backend-independent contract, readiness, simulation, preparation, specification, and prior-predictive functionality.
• Added restricted binary model translation from approved package specifications to brms Bernoulli-logit formulas and priors.
• Added optional full-MCMC fitting through the fixed brms and rstan sampling route without unrestricted formulas or backend arguments.
• Added conservative fit metadata that records sampling settings while explicitly withholding convergence and posterior-adequacy claims.
• Added conservative binary posterior diagnostics covering R-hat, bulk and tail ESS, divergences, maximum-treedepth saturation, and chain-level energy diagnostics.
• Added posterior summaries, binary posterior predictive checks, prior-scale sensitivity, simulation-based recovery, diagnostic plots, and structured Markdown model reports.
• Diagnostic, predictive, sensitivity, and recovery statuses never create automatic convergence, adequacy, robustness, or validation claims.
• Added the complete hierarchical lognormal duration workflow for strictly positive finite uncensored outcomes.
• Added deterministic duration simulation, explicit unit conversion and preparation, model specification, prior predictive checks, restricted brms translation, and full MCMC fitting through rstan.
• Zero, negative, censored, truncated, shifted, survival, Gamma, Weibull, and mixture outcomes remain outside the approved duration contract.
• Added conservative duration posterior diagnostics, posterior summaries with median-ratio interpretation, posterior predictive checks, prior-scale sensitivity, simulation-based recovery, and structured Markdown reports.
• Duration validation statuses remain separate from automatic convergence, adequacy, robustness, causal, or substantive claims.
• Added integrated end-to-end binary and duration vignettes.
• Added dedicated articles for sampling diagnostics, prior sensitivity and recovery, and optional backend installation.
• Added a repository-only audit covering exports, Rd aliases, pkgdown reference topics, articles, built pages, and optional dependencies.
• Aligned DESCRIPTION, citation metadata, README, package documentation, and the curated pkgdown reference and article indices with the complete scope.