IJSE 0.3.4

Accept an S-by-1 pointwise likelihood matrix from categorical brms models, preserving draw and observation order; reject transposed or multi-column results.
Reject covariance ARMA extraction, whose brms output contains LOO conditional densities rather than joint likelihood factors. Explicit joint sampling-unit matrices remain supported through ijse_draws().
Require explicit likelihood matrices for discrete truncated models after detecting a lower-bound normalization discrepancy in brms 2.23.0.
Reject measurement-error prediction with unsaved latent draws, which brms otherwise resamples. Document conditional versus observed-data likelihoods and missing-response NA rejection.
Add real fitted-model regressions for the new streaming support and guards.

IJSE 0.3.3

Reject missing cluster and chain IDs encoded as a used explicit NA factor level.
Check labeled likelihood blocks against one another when draw rows are unlabeled.
Avoid an unnecessary first-block mean square that can overflow despite finite centered sampling variance.
Add regressions and actual weighted Bernoulli/missing-row and nonlinear distributional Gaussian fixtures with direct density checks and draw/block parity.
Add scale-sensitive independent algebra, joint nonlinear reweighting, metadata, posterior-format and targeted fault-injection audit tools.
Clarify the plain-matrix transform contract and cross-block row-label validation.
Load study test helpers before test execution so shuffled runs are independent of top-level expression order.

Preserve named-vector transform row identities and reject detectable draw permutations instead of silently computing covariance against the wrong rows.
Stabilize normal intervals at confidence levels close to one or zero; keep column labels distinct when percentages coincide at display precision.
Make display labels unique and retain exact original cluster/chain IDs.
Repair Gaussian/quantile fixture clustering checks: sampling IDs are saved separately and tests require eight actual clusters rather than accepting NULL.
Add boundary regressions and preserve correct automatic data-frame row handling.
Add a package overview, explicit return-field definitions, normalization and alignment rules, published references, and executable no-refit brms examples.
Add a renderer for HTML and PDF reference manuals from the actual Rd sources.

Expand the introductory vignette and add executable brms and diagnostic tutorials.
Include two newly fitted synthetic median-regression examples with fitting data and sampling-unit IDs saved separately, plus generation seeds and diagnostics.
Install standalone R scripts extracted from the tutorial sources, with fresh process verification and a collected PDF/HTML rendering workflow.
Add regression checks for tutorial fit/row/cluster correspondence and an independent numerical audit of covariance, transformations, blocking and permutation/translation invariance. The numerical R API is unchanged.

Fix chain metadata for posterior data frames with unequal retained chain lengths.
Add all/posterior/none diagnostic modes and skip discarded diagnostics in IJ_se().
Retain posterior covariance separately; add vcov(type = “posterior”) and a standard influence() accessor for explicitly saved influence values.
Avoid redundant centering and copying of the selected quantity draws.
Add a reproducible vignette study with 40,000 exact-posterior simulations, 320 independent posterior-draw audits and 400 actual brms median-regression fits.
Include study generators, compact results, Monte Carlo uncertainty, fit-attempt accounting and figures. Study simulations are not run during package checks.
Add regression tests for chain metadata, diagnostic modes and study calculations.

Reject missing cluster IDs instead of silently omitting observations and counting an empty cluster. Reject undefined variances and non-finite inputs.
Preserve likelihood extraction errors instead of replacing every error with generic save_pars advice.
Keep IJ_se(fit, cluster_var) as a named-vector compatibility interface.
Add ijse() and ijse_draws(), full covariance, transformed quantities, explicit parameter/draw selection, optional influence values, and standard S3 methods.
Add callback likelihood blocks and univariate brms streaming. Aggregate cluster covariance contributions without allocating a draws-by-clusters matrix.
Report chain stability checks and quantity-level posterior diagnostics.
Make brms optional for the backend-independent matrix interface.
Replace placeholder tests with numerical, analytic, and real brms regression tests. Add a vignette, README, reproducible fixture generator, and benchmark.
Generate explicit exports and correct S3 registration from roxygen source.
Add the published method citation. Document estimands and inference limits.

Added a NEWS.md file to track changes to the package.
Updated example code to include cases for quantile regression on both independent and clustered data.
Optimized code structure