Changelog • influ2

influ2 1.1.0

Clarified the distinction between numerical verification and scientific calibration in residual diagnostics, and the nominal, conditional meaning of implied-effect intervals. Conditioning tables now include lme4. Domain documentation makes declared support, within-cell area averaging, and non-causal term-omission sensitivity explicit. Calculations, defaults, and frozen validation results are unchanged.
Added index_domain() for explicit sampled-domain coverage, link-scale decomposition, sampled-only indices, and prediction-omission counterfactuals. Stored influ_domain results have table, summary, print, and plotting methods, including stacked coverage and response partitions, full-index comparisons, paired-draw combined unsampled shares, omission ratios, retained-cell maps, and a sampled-only comparison with a coverage panel. Area totals are individually normalised over common years to compare trend shapes. These diagnostics do not refit reduced models or turn a changing sampled domain into a fixed-domain abundance index. Sampled-only results compare all three support definitions; omission uses one definition ("cell_time" by default). Omitting a sampling rule uses half the reference-grid spacing with a warning, not an inferred footprint. cpue_index() also supports term-omission counterfactuals through omit_terms, omit_where, and explicit sampling/reference rules. Domain diagnostics and omissions also support single-predictor mgcv GAMs, using native smooth contributions and paired joint Gaussian coefficient draws. Spatial and spatiotemporal smooth grouping is explicit. Executed GAM, sdmTMB, and tinyVAST examples demonstrate the workflow.
Domain coverage now excludes missing-response and zero-likelihood-weight fitted rows, default plots handle empty always-sampled cores explicitly, and restricted integrated indices retain their actual areas rather than inheriting full-domain areas. Truncated negative-binomial native response predictions now use each draw’s dispersion, including delta positive components. Saved domain summaries and plots are restart- and locale-stable without loading the fitting backends.
Added plot_cdi() for one or more model runs in one shared CDI display. Coefficients and composition bubbles use consistent model offsets and colours, and annual influence lines share the year axis. Unequal year ranges, unavailable categories, and internal gaps remain explicit. Saved summaries retain their original centring and uncertainty; plotting does not refit models or silently impose a common reference.
Added a separate index_ratio() diagnostic and plot_index_ratio() for comparing calculated indices, standardised and nominal series, and stored model steps. Ratios use matching years, a named denominator, and optional common-period display normalisation. Undefined denominators remain gaps; point ratios do not invent intervals from marginal model uncertainty.
Added search_cpue() and compact influ_search reports for constrained, buildmer-style forward, backward, bidirectional, and order-then-backward searches using seven native model backends. Annual blocks, offsets, fitted rows, and essential processes stay protected; non-essential random terms can be explicitly nominated. Preflight is the default, native refitting requires refit = TRUE, candidate evaluations are cached and bounded, and selected_model() retrieves the best eligible evaluated fit. Native numerical/criterion failures remain visible, and no residual p-value selection, automatic process toggling, or model averaging is implied.
The CPUE-indices article now demonstrates an area-weighted lobster CPUE mean, equal-habitat weighting, native joint prediction/covariance checks, and the distinct area-integrated CPUE total. The response stays in CPUE units; no abundance or biomass conversion is implied.
Compact DHARMa residual calculations now use temporary disk-backed joint simulations and bounded observation blocks by default. Native DHARMa PIT and vector-based tests are retained; dispersion and outlier tests use explicitly labelled bounded-memory equivalents. dharma_storage selects automatic, disk, or full-memory storage, and native-object retention keeps its complete matrix in memory. Temporary caches are removed on success or failure. The residual article compares disk and native-memory results. Equivalent tests are validated for DHARMa 0.4.7 and 0.5.0; other versions leave those tests explicitly unassessed, with native memory-mode tools available separately. Simulation conditioning and fitted-model calibration are unchanged; the workspace budget is not a total-process memory guarantee.
index_table() labels interval columns with their stored probabilities, such as Q0.025/Q0.975 or Q0.05/Q0.95, without changing any values. quantile_names = "generic" and as.data.frame() retain fixed Qlower/Qupper names. Year-effect indices retain their diagnostic’s actual interval probabilities, including when extracted from saved diagnostics.
Added check_cpue() with compact overall native model statistics, numerical fitting evidence, and an explicit inventory/check for every model term across seven backends. Raw-predictor PIT association screens are exploratory, with Holm adjustment, not coefficient-significance or term-removal rules. Missing latent, interaction, component, and sparse-data checks are not passes. Saved native TMB objectives are not reinitialised implicitly. Printing and plotting use retained results, without model refits, MCMC, or new simulations.
Optional residual_tests = "standard" or "full" calculates genuine native DHARMa statistics before discarding its response matrix. annotations adds stored statistics to residual plots, with sources, conditioning, and finite Monte Carlo precision explicit. Existing residual and plot defaults remain unchanged. A worked mixed-lobster report documents overall/term interpretation and the correlated-residual false-positive warning highlighted by Andrea Havron and Cole Monnahan; calibrated acceptance tests and OSA remain separate.
Added optional lme4 adapters for lmer, glmer, and glmer.nb fits across influence/CDI diagnostics, PIT/DHARMa residuals, likelihood-implied effects, CPUE indices, area integration, annual covariance, comparisons, and step plots. Gaussian log-response interpretation is explicit. Joint coefficient draws preserve fixed/random dependence; variance and dispersion parameters are held at native fitted estimates. Negative-binomial step refits require an explicit choice about re-estimating size. REML criteria are not ranked across fixed-effect structures. A worked simulated-lobster example is included.
Added optional DHARMa residual calculation to influ_residuals() and as_influ_residuals(), reusing native response simulations, conditioning, and component selection. DHARMa owns the PIT calculation; uniform endpoints are preserved, with explicit finite limits in normal-score displays. Results stay compact by default; retain_dharma = TRUE keeps the genuine DHARMa object for its own tools. Full-matrix allocation has a configurable memory guard. The default residual engine and likelihood-implied effects are unchanged.
Expanded residual-implied effects to ordinary binomial GLM/GAM/glmmTMB fits, sdmTMB Gaussian/Gamma/Poisson/NB2, and native Tweedie fits across the five frequentist backends. A verified brms custom Tweedie contract is also supported. Added standard spatial delta-Gamma, glmmTMB Gamma and truncated-count hurdles, brms Gamma/count hurdles, and zero-inflated Poisson/NB2 mixtures (plus glmmTMB zero-inflated Tweedie). Mixtures retain zeros in their full likelihood; extra-zero effects are distinct from observed encounters. Native nuisance parameters remain fixed, and boundary/unidentified shifts remain explicit.
CDI plots accept coefficient_order = "ascending" or "descending" to order categories by their displayed effects. Coefficient and distribution panels stay aligned, without changing diagnostic values or annual influence. The default, "original", preserves the existing plots.
Sped up the full lobster lme4 vignette fit using a validated native optimiser, without changing its data, model, or likelihood integration. Development checks limit numerical-library threads and now include the PDF manual.
CDI encounter panels use the compact label “Effect (log-odds)” while positive ratio panels retain “Relative Effect”. Shared fixed/random-effect plots use horizontal short category labels. Matching upper/lower labels are spaced automatically from the rendered text and panel width: all when they fit, otherwise every second, third, or subsequent category. There is no fixed label-count limit, and no plotted levels or ticks are removed. This uses the public axis-guide extension API in ggplot2 >= 3.5.0.
Joint sdmTMB step plots now support paired delta-Gamma/lognormal formulas, explicit annual predictors in each component, and encounter, positive, or combined fixed-year contrasts. Reduced stages preserve fitting rows, effort offsets, weights, and the spatial mesh; unchanged full stages are reused. Joint fixed-effect covariance is retained for combined uncertainty. These are not area-integrated indices or full offset-dependent CDI decompositions.
Residual-implied effects now include tinyVAST and brms adapters for Gaussian, Poisson, NB2, Gamma, Bernoulli, and lognormal responses with validated links. Standard tinyVAST delta-lognormal and brms hurdle-lognormal fits expose explicit encounter, positive, and combined views. Native spatial fields, random effects, smooths, offsets, and dispersion are retained. brms defaults to fixed posterior-mean parameters; draw_id selects one joint posterior state. Labels explicitly identify conditional likelihood intervals, not Bayesian credible intervals. Native-likelihood tests use small saved genuine posterior fixtures; no MCMC is run by diagnostics, tests, or vignettes.
Added explicit encounter, positive, and combined residual-implied diagnostics for standard delta-lognormal sdmTMB fits, plus single-response Bernoulli(logit) and lognormal(log) support. Native offsets, vessel effects, spatial fields, spatiotemporal fields, and dispersion remain fixed. An explicit year_term handles components with different annual specifications. Combined displays show original-observation expected responses, not implied coefficients or standardised indices; conditional profile intervals allow both component adjustments. Sparse, empty, and boundary strata remain explicit.
Added worked omitted-covariate and year-specific spatial residual examples to the Residual diagnostics article, reusing saved N09 results without model fitting or response simulation. Shared scales, guarded row matching, and explicit conditioning illustrate patterns that pooled checks can miss. The diagnostic defaults and frozen validation results are unchanged.
Added a worked Regional CPUE indices article (#18): one negative-binomial glmmTMB model, four areas, and two fixed-reference regional series, using the existing index API. It demonstrates interaction-supported trends, known simulated truth, weight sensitivity, relative normalisation, and assessment tables/covariance, with independent native-prediction checks. Separate regional results do not supply cross-region covariance; the article explicitly guards against interpreting them as independent.
Added index_table() for assessment reporting, including explicit moment-matched Meanlog, SDlog, and LognormalMedian columns. Reporting omits an unavailable median by default, while as.data.frame() preserves the existing schema and genuine posterior medians remain unchanged (#22).
Standardised and area-integrated indices now retain compact response- and log-scale annual covariance matrices across all six supported backends. index_vcov() and vcov() extract them with explicit year alignment; plot(index, type = "covariance") or "correlation" displays them without recalculation. Preview/old objects and undefined log covariance fail clearly. Relative-index singularity is preserved, not repaired with jitter. Worked assessment examples distinguish joint covariance from marginal lognormal approximations, model-coefficient covariance, and additional assessment error.
plot_data_extent() now leaves zero-completeness cells blank, without bubbles, while retaining all variables and years on the axes. Red background markers have also been removed from positive-completeness cells: dark bubble area alone shows the proportion present. Zero-valued observations still count as present. The main article and help describe this distinction.
Added Langley’s (2018) SNA 7 assessment to the residual-implied effects article, with pointers to its depth, month, target-species, and vessel comparisons and an explicit distinction from the new likelihood method.
Added Gamma(log) residual-implied effects for GLMs, mgcv GAMs, and ML glmmTMB fits, using native fitted dispersion and a stable analytic log-mean adjustment. Conditional profile intervals are checked against independent Gamma densities, including varying dispersion. Smooths, fitted random effects, effort slopes, and offsets remain fixed; no model or Gamma shape is refitted. Annual trajectories now also break across numeric years wholly absent from the fitted factor levels.
Added IV01, a frozen 400-fit NB2 glmmTMB validation of residual-implied effects, with balanced/uneven sampling, null/omitted-trend cases, and known-parameter controls. The article separates injected interactions from fitted-conditional targets and reports interval limitations. Package tests audit compact results without rerunning the study. No diagnostic algorithm, family support, or default changed.
Corrected the optional PIT ECDF and difference plots’ reference-grid mismatch identified by N09. bayesplot’s independent simultaneous limits are retained, aligned with the empirical CDF, and completed at zero. pit_grid_size = K now displays all K + 1 points of (0:K) / K. Endpoint masses are retained; stored residuals and default panels are unchanged. Regression tests cover the frozen N09 limits, ties, endpoints, and future dependency-layout guards. The independent-uniform control now reproduces 4.77% crossings, versus 12.46% before the fix; this is not a fitted-model calibration claim.
Restored effect-scale plot_implied_residuals() with a compact implied_effects() calculation and likelihood-based local adjustments. Initial adapters cover Gaussian/log-response, Poisson, and NB2 models in lm, GLM, GAM, and glmmTMB. Explicit traditional options distinguish ordinary log-response arithmetic from the historical standardised-residual GLM convention. The new article verifies both comparisons. Conditional intervals are not full interaction/index uncertainty; unsupported cases fail clearly. The previous grouped PIT calculation is unchanged under its clearer name, plot_grouped_residuals(); update callers that pass stored PIT objects.
Added a frozen, reproducible N09 residual-validation study: 100 NB2 datasets each for glmmTMB and sdmTMB, known-truth controls, omitted-structure cases, and seed/count sensitivity. The article reports conditioning-specific behaviour and an independent bayesplot 1.16.0 PIT-band grid-alignment finding. No PIT calculation or conditioning default changed. Monte Carlo fits are not rerun by CI or vignette builds.
Residual and index calculations now name a missing optional model backend before trying to extract its formula or family. Saved compact results remain usable without those backends. An isolated minimal-installation smoke check now runs within the existing Ubuntu-release CI job.
Added clean-session save/reload regression tests for compact influence, residual, CPUE-index, and step-comparison results, including mixed-model and Bayesian examples. The main vignette shows how to save the results and reopen their tables and plots without repeating model calculations.
Added a worked plot_data_extent() example near the start of the main vignette and in the helper’s documentation. A separate lobster-data copy illustrates missing measurements without changing any fitted examples; zero responses remain observed values, not missing data.
Shortened the ratio-scale CDI y-axis label to “Relative Effect” for all terms. Reorganised the website reference and article lists, with residual diagnostics first and Bentley validation last. Calculations are unchanged.
Added explicit residual conditioning options. Existing defaults remain: fitted effects for GLM/GAM/sdmTMB/tinyVAST, new random effects for glmmTMB, and posterior predictive draws for brms. New options hold glmmTMB effects fixed, reuse one joint conditional latent-effect draw for sdmTMB/tinyVAST, or simulate new sdmTMB processes. Unsupported combinations fail explicitly. Independent native objectives protect the supplied fit. Shared draws and per-replicate seeds preserve the new schemes across response batches; sampled-field encounter probabilities and delta components stay aligned. The spatial article compares schemes on the same fitted models. These are exploratory diagnostic targets, not interchangeable calibrated tests.
Added as_influ_residuals() for externally generated response simulations. It requires explicit observation alignment, response kind, time column, and simulation conditioning, and returns the same compact diagnostic used by the standard plots and residual helpers. Bernoulli/binomial calibration requires original fitted probabilities and known trial counts. Combined and component inputs are never silently converted or filtered.
Extracted one shared residual-summary engine, protected by a frozen pre-refactor numerical/RNG baseline and native replay tests. Existing fitted model defaults and calculations are unchanged. The supplied-matrix route does not simulate, refit, retain the full matrix, or add a streaming interface. The residual article demonstrates it with the fitted lobster glmmTMB model.
Added optional standalone PIT ECDF and ECDF-difference panels through bayesplot, reusing stored PIT values without additional response simulation. Their simultaneous iid-uniform reference limits are labelled as exploratory, not calibrated tests for fitted models. panels now selects and orders any four supported residual panels; the default layout is unchanged.
Expanded website figures now display the complete numbered caption from beneath each figure, rather than its shorter accessibility description. Long captions remain accessible, and the website build checks every plotted image for a matching caption. Calibration plots retain their separate grey predictive bars; the residual article clarifies their interpretation.
Labelled the four-panel overview’s normal-score PIT residuals explicitly in its footer and residual axes. The footer distinguishes panels A-C from the response ECDF or probability-calibration check in panel D. Help and the residual article explain the existing qnorm(pit) transformation; residual calculations, simulation schemes, and default panel choices are unchanged.
Added standalone lobster ECDF and complete-fit brms posterior predictive examples to the residual article, with compact saved Bayesian results and an individual-replicate overlay. Help and tests cover standalone/overview ECDF agreement and calculation reuse. Fitted-data ranks, response ECDFs, LOO-PIT, and LOOIC are explicitly distinguished; no universal LOO-PIT or MCMC-on-render workflow is introduced.
Unified the remaining residual plotting helpers with influ_residuals(). plot_predicted_residuals() now uses normal-score rank residuals and the predictive means from the same simulations. plot_implied_residuals() now shows year-by-group mean generalised departures around zero, not residuals added to coefficients on a different scale. Explicit native residual types are rejected with a migration message. Both helpers accept calculated residual objects, preserve response-component metadata, and avoid repeated simulation when restyling. Retain grouping columns with groups during calculation. Historical source and frozen review figures are unchanged. The Model comparison article now includes an executed six-backend table including a complete brms fit, without running MCMC during rendering.
Extended table_criterion() to GLM, GAM, glmmTMB, sdmTMB, tinyVAST, and complete brms fits, including mixed-backend lists. Native AIC, BIC, deviance, degrees of freedom, and Bayesian summaries use separate columns. Explicit conditional AIC uses mgcv, sdmTMB, and tinyVAST native methods; unsupported glmmTMB mixed-model cAIC stays missing with a reason. Input order is now preserved by default. Optional differences and sorting require compatible fitted observations and likelihood targets; REML, penalised, failed, and unreliable-LOO cases are not automatically ranked. Paired LOO differences support both older matrix and newer data-frame outputs from loo. The new Model comparison article includes an executed compact Bayesian/frequentist table; no MCMC is run during rendering.
Retired plot_qq() without a compatibility wrapper. Calculate checks <- influ_residuals(fit), then use plot(checks, type = "qq") for the same simulation-based Q-Q diagnostic as the four-panel overview. The residual article includes a standalone glmmTMB example, and the plotting help cross-references this workflow. Its pointwise reference ribbon is not posterior uncertainty around individual points. Native-residual Q-Q plots and per-point posterior intervals are not part of the first-release API. The frozen legacy article remains available for review.
Fixed glmmTMB influence diagnostics with explicit reference grids: the fitted model-matrix method ignores newdata, so fixed-effect designs now come from the native prediction setup. Regression tests cover different and equal-sized reference grids, polynomial bases, zero-inflation terms, and random-effect alignment. cpue_index() uses a separate prediction path and is unchanged.
Fixed reference-predictor checks for brms distributional terms without random effects, where partial field matching could interrupt CPUE-index calculations. Added numerical and contract tests for index comparisons, joint uncertainty, prediction inputs, and refitted step sequences.
Standardised expected-response CPUE tables now include sdmTMB and univariate tinyVAST, alongside GLM, GAM, glmmTMB, and complete brms fits. Spatial indices use native combined response predictions and compact joint Gaussian parameter/field uncertainty, with explicit field and exposure choices. Shared draw identities are invariant to prediction and draw batch sizes.
Added integrate_index() for all six model types, with or without spatial terms. It sums expected responses times supplied cell areas and records response/area units, known catchability conversions, and optional within-cell seasonal averaging weights. Standardised means and integrated totals remain distinct calculated objects for plotting and comparison. The CPUE and spatial articles demonstrate these workflows; frequentist medians and the remaining legacy-function review are unchanged and explicitly deferred.
Residual overviews now select encounter probability calibration for Bernoulli responses, while retaining ECDFs for positive, combined delta/hurdle, count, and grouped-binomial responses. Fixed fitted-probability bins preserve ties; compact native simulation summaries supply pointwise predictive envelopes. Added explicit calibration and scientific grouped checks, known-trial guards, component-aware hurdle/SDM simulation routing, and informative legacy-object fallbacks. Existing residual values and explicit distribution panels are unchanged. The residual article demonstrates good, distorted, and deceptively pooled calibration. No model fitting or cross-validation is triggered by plotting.
Restored geo_mean() as a stable, documented public utility. Added cpue_index() assessment tables and plot_index(), with both standardised and standardized spellings accepted. Expected-response indices remain distinct from existing year-effect contrasts. Response standardisation initially supported GLM, GAM, glmmTMB, and complete brms fits, with joint delta-method or batched posterior uncertainty and explicit reference-population choices. plot_compare() now also accepts calculated index objects without changing its existing fitted-model behaviour. The CPUE indices article documents the table schema, uncertainty meaning, and model-specific prediction choices.
Year-panel residual box widths now encode the square root of sample size; year labels no longer contain sample counts. Corrected the package name to lowercase brms throughout maintained documentation and messages.
CDI proportion legends use one column with at most four reference bubbles, preserving the composition panel’s bubble-size mapping. The spatial article hides the mesh-margin setup and restoration while still executing both.
Refined CDI layouts: short term labels are horizontal and repeated above the fitted-effect panel; focus/year labels sit on the influence panel’s right-hand axis. The two-column size legend is extracted without unused guide slots, preventing its title from being clipped. Four-panel residual examples now use glmmTMB models with a monthly random intercept. Bentley comparison labelling and the native mesh figure’s margins are tidier.
Added influ_residuals() and a reusable four-panel residual display: simulation-based Q-Q, residuals against the predictive mean, residuals by automatically detected fishing year, and observed versus simulated ECDFs. Complete response simulations are processed in small batches; fitted models and observation-by-simulation matrices are not retained. Backend conditioning, finite-rank randomisation, and reference-band limitations are explicit. An explicit year override handles ambiguous time terms, and numeric years retain gaps in sampling. The residual article demonstrates full and misspecified lobster models. The frozen legacy review article is preserved.
Completed the first legacy-helper triage: retired plot_hurdle(), get_coefs(), get_coefs_raw(), get_marginal(), and the old Shiny influ_app(). No retained implementation depends on these functions. Consolidated get_influ(), get_influ2(), and plot_influ() on the model-neutral influence interface, and plot_bayesian_cdi() and plot_bayesian_cdi2() on plot(..., type = "cdi"). Their frozen source and help files were removed without adding compatibility wrappers. The current influ() API, hurdle/zero-inflated support, frozen review page and figures, and undecided assessment-output helpers remain intact.
Added a residual-diagnostics article and reference section, with executable lobster examples, optional DHARMa checks, and guidance on PIT, spatial, Bayesian, and one-step-ahead residuals.
Aligned residual plots with fitted observation rows after omissions, subsets, and reordered data. Ambiguous multivariate, delta, and compact brms inputs now fail informatively instead of producing mismatched diagnostics.
Preserved negative additive contrasts in model comparisons, rejected mixed index scales, and prevented duplicate labels from merging model curves.
Added the model-neutral influ() S3 API and compact influ_diag result.
Added adapters for GLMs, mgcv GAMs, brms, glmmTMB, sdmTMB, and tinyVAST.
Separated uncertainty calculation from draw retention, including a posterior-mean preview mode and derived-draw or disk retention.
Added common print, summary, plot, and autoplot methods.
Added influ_steps() for ordered comparisons of genuinely refitted models, with automatic main-formula sequences for simple GLMs, GAMs, and glmmTMB models, explicit step specifications, and reuse of already fitted models. plot_step() accepts the same refitting route or a reusable influ_steps result. Step plots compare centred year-effect contrasts, with 95% intervals by default; they do not calculate area-weighted abundance indices.
Added negative-binomial GLM refitting through MASS::glm.nb(), preserving formula offsets, weights, and analysis rows while re-estimating dispersion at each changed stage. The main lobster step demonstration now uses this model and its approximate negative-binomial confidence intervals.
Guarded ambiguous interaction-only focus indices and implied-residual baselines. Reference grids centre effects; they do not automatically marginalise interactions.
Restricted offset/exposure diagnostics to supported single-component log-link ratios and identity-link contrasts. Other links and combined hurdle/zero-inflated calculations with offsets fail explicitly. Nominal summaries remain observed-response means, not exposure-adjusted CPUE.
Added backend-specific lognormal guards: brms requires constant log-scale sigma and an identity location link; mean-parameterised backends require log links. glmmTMB log-mean ratios remain available with varying dispersion.
Centred CDI fitted-effect panels on the same weighted reference as influence. Log-link and lognormal components now display relative effects about one on a logarithmic axis; other links retain clearly labelled additive units. Confidence and credible intervals propagate uncertainty in the estimated reference, and posterior ratios are summarised after transforming each centred joint draw. The previous model-coded link display remains available with coefficient_reference = "model".
Added the initial Gaussian, binomial, Poisson, negative-binomial, lognormal, Gamma, and Tweedie family registry. Quasi and specialist extended families are deliberately excluded.
Moved the original proto implementation out of the runtime namespace and added a frozen Bentley parity fixture and tests.
Replaced the introductory and hurdle vignettes with model-neutral design and reference documentation.
Added sparse joint-precision uncertainty for spatial and spatiotemporal fields, including draw-by-draw delta-field combinations.
Added explicit prediction-grid and reference-weight standardisation, multivariate mixed-family tinyVAST responses, CDI coefficient intervals, and model-neutral Bentley overall and trend metrics.
Propagated the joint conditional latent covariance for glmmTMB random effects instead of treating conditional modes as uncertainty-free.
Retained selected comparison, model-criterion, data-extent, residual, Q-Q, and step-plot helpers, while moving their brms and rstan tooling out of mandatory dependencies. The remaining historical helper source is frozen under tools/legacy/ for explicit function-by-function triage.
Replaced embedded fitted-model test data with compact posterior-draw fixtures, substantially reducing the source-package size.
Replaced the lobster example with a reproducible, simulated dataset and added stricter validation for intervals, weights, focus ordering, component draws, data-extent plots, bubble plots, and implied residuals.
Strengthened the simulated lobster teaching scenario with changes in sampled season, depth, and soak time. Retained its known annual effects as dataset metadata and added a dynamically calculated truth check for the refitted negative-binomial GLM sequence. The brms example includes soak time alongside monthly and depth effects. Documentation distinguishes point-estimate recovery from interval coverage and explains the model-based intervals.
Corrected brms population-term mapping when ordinary predictors follow group-level or smooth terms, so their labels and CDI grouping match the fixed-effect design matrix.