Changes in version 0.5.1 (2026-09-30) - Fix WAIC portability on macOS ARM and R builds without extended-precision long double arithmetic. Center and scale log-likelihood contributions before computing their dispersion, avoiding overflow in intermediate sums. - Use the same stable dispersion calculation for paired information-criterion standard errors. Return an explicitly unavailable WAIC when finite inputs produce pointwise or aggregate scores outside the representable range. - Retain the extreme-value regression that exposed the problem, and add tests for double-only arithmetic, large constant offsets, genuine overflow and agreement with ordinary-range WAIC calculations. Priors and posterior sampling code are unchanged from 0.5.0. Changes in version 0.5.0 (2026-09-21) - Harden the pre-release numerical audit: use exceptional-range log-scale fallbacks for the weighted Lindley mean inversion and censored transforms; avoid rounding a positive beta shape to zero; stabilize SD and mean ESS under changes of units; flag non-moving MCMC chains explicitly. - Handle sure censoring events at zero analytically in PSIS-LOO instead of fitting degenerate Pareto tails. Guard aggregate criterion/SE overflow, withhold DIC when the likelihood at the posterior mean is undefined, and retain completed posterior draws if optional post-processing fails. - Add permanent regression tests for these audit findings. Priors and sampling algorithms are unchanged; ordinary-range regression outputs remain identical for all 56 complete-data model-prior routes. - Reject unused built-in fitting arguments and non-vector/complex data rather than silently ignoring arguments or flattening arrays. Do not clip boundary Beta draws in the complete-data geometric posterior sampler. - Add opt-in criteria = FALSE to both fitting interfaces. TRUE computes WAIC, PSIS-LOO, and DIC after sampling; character vectors select a subset. Add criteria(), WAIC(), LOOIC(), DIC(), and compare_models(). Default fitting does not evaluate additional likelihoods or load 'loo'. - Use the observed likelihood under censoring, guard leave-one-out posterior propriety, respect certified parameter moments for DIC, and report pointwise diagnostics and paired comparison uncertainty. PSIS-LOO uses the optional 'loo' package; WAIC and DIC need no additional dependency. - Add the separate function fitcensBayes(x, status, distr, prior) for independent right censoring, with status 1 for exact events and status 0 for strictly greater lifetimes, including discrete distributions. - Support the 20 registered distributions and 56 complete-data model-prior routes. These priors are reused, not rederived for a censoring design. Check sufficient posterior-propriety conditions on the exact-event subset before sampling; handle the Exponential posterior analytically. - Add exact censored Gamma and Beta posteriors where available, collapsed Weibull sampling, observed-likelihood slice sampling, and optional data augmentation with truncated lifetimes. Improve numerical evaluation of extreme Rician tails in the new censored interface. - Add the fitcensBayes class, posterior summaries and diagnostics, survival prediction, imputation, and observed pointwise log-likelihood extraction. Only certified posterior moments are reported. - Add fitcensBayes_models(), help pages, examples and regression tests. - Preserve sampling algorithms and ordinary-range regression results; the new criteria argument is named-only to preserve existing positional calls. Changes in version 0.2.3 (2026-09-09) - Reject tied observations for the Student-t model with unknown degrees of freedom under the independence Jeffreys prior. Such samples have an improper posterior because the prior gives positive mass to degrees of freedom for which the location-scale posterior is not integrable near zero scale. - Document the pairwise-distinct sample requirement in the help page and public catalogue, and add regression tests for tied and distinct observations. - The fixed-degree-of-freedom Student-t routes retain their existing, multiplicity-dependent propriety conditions. Sampling algorithms and the remaining model-prior routes are unchanged. Changes in version 0.2.2 (2026-08-29) - Added the weighted Lindley distribution with Fisher-information Jeffreys, one-group reference, Jeffreys' first-rule, independence-Jeffreys, and exact ordered reference priors for lambda and phi. All six routes enforce the proved condition that the sample is positive, nonconstant, and has at least two observations. - Disabled weighted Lindley MDI fitting because the corresponding posterior is improper for every finite sample. - Added the closed-form weighted Lindley likelihood initialization with a numerical maximum-likelihood fallback, stable log-prior evaluation, a Fisher-orthogonal mean/shape Metropolis parameterization, mixture-based posterior prediction, and pointwise log-likelihood evaluation. - Expanded the public catalogue to 20 distributions and 56 enabled model--prior routes, with analytic, initialization, predictive, guard, and registry tests for the new family. Changes in version 0.2.1 - Added fitdistrBayes_model(), a public model specification for combining a user-defined density and prior with parameter support, starting values, validation, posterior-propriety and moment declarations, posterior prediction, and an adaptive Metropolis, slice, or user-supplied sampler. - Extended fitdistrBayes() so the distribution argument can be a fitdistrBayes_model object. The fitted object records which extension components and computational engine were used. - Added executable tests for a Laplace model with a proper non-objective prior, a slice-sampled Exponential model, and an exact Poisson--Gamma posterior simulator. Changes in version 0.2.0 (2026-08-06) - Expanded the help pages to document the class, structure, and statistical meaning of every exported function and S3 method return value. - Replaced the non-executable tutorial wrapper with an interactive file-view example and left all fitting examples unwrapped after verifying that the complete help example runs in less than five seconds. - Removed explicit access to .GlobalEnv and all superassignment operators from the package code and validation scripts. - Added Gumbel, Frechet, Lomax, Nakagami-m, Exponential-Logarithmic, and Rician models with documented propriety conditions enforced on the supported data domain. - Added closed-form moment starts for Gumbel, Nakagami-m, and Rician, closed-form L-moment starts for Frechet and Lomax, and a stable one-dimensional moment start for the Exponential-Logarithmic model. - Added marginal/conditional samplers for Frechet and Nakagami-m, stable transformed Metropolis samplers for the other new models, and an accelerated Gauss-Laguerre evaluation of the Rician Jeffreys factor. - Reparameterized the Lomax Metropolis sampler by shape and scale/shape to reduce posterior correlation while retaining the documented output scale. - Added explicit failures for the improper Lomax reference posterior and Nakagami-m MDI posterior, as well as for unproved Rician reference/MDI routes. - Expanded the analytic, numerical, initialization, route, and stress tests to 19 built-in distributions and 50 enabled model-prior routes. - Added fitdistrBayes_routes() as a machine-readable catalogue of all 50 enabled routes and made the route test execute directly from this catalogue. - Expanded help("fitdistrBayes") with all built-in models, parameterizations, priors, fixed-parameter requirements, propriety guidance, and one example for every family; the English-language 50-route console tutorial is installed as examples/tutorial_fitdistrBayes_all_models.R. - Tightened the Nakagami-m propriety check to reject constant positive samples. - Made the custom density/prior log-mode contract explicit and reject undefined or positive-infinite log values; custom RNG outputs can now be validated. - Reported post-warmup, warmup, and all-iteration acceptance rates separately, included tail ESS in the convergence decision, and corrected constant-chain and per-chain ACF diagnostics. - Removed silent floating-point clipping: unrepresentable natural-scale draws now produce transparent numerical-range errors. Changes in version 0.1.2 - Added classical automatic initialization, including the closed-form Weibull L-moment estimator. - Made non-constant-sample validation invariant to translation. - Replaced costly truncated negative-binomial entropy sums with an exact numerical integral for MDI fitting. - Stabilized natural-scale Gamma and Weibull draws at floating-point limits. - Preserved the matrix shape of log_lik() for one-observation samples. - Scaled Metropolis location proposals to the model's initial scale, fixing poor mixing for very small- and very large-scale data. - Added broad stress, entropy, initialization, edge-case, and difficult-chain regression tests.