Changes in version 0.2.3 (2026-10-28)                  

  - Shorten the package title and add a standard loss-modeling reference
    in DESCRIPTION.
  - Rewrite help-page return-value descriptions and important argument
    explanations.
  - Clarify ground-up observation censoring versus post-deductible
    payment caps.
  - Document deterministic analytic plotting as the default and
    simulation as optional.
  - Add installation instructions and a reproducible release-check
    procedure.
  - No changes to likelihood, optimization, plotting, or
    factor-calculation code.

                        Changes in version 0.2.2                        

  - Fit a zero-intercept nested anchor before estimating a free scaling
    intercept. Reject an apparently converged free-intercept fit if its
    log likelihood is materially worse than the nested solution.
  - Retry Gamma/Burr optimizations from multiple shape starts, report
    optimizer_attempts, reject inadmissible final fits, and avoid
    pathological near-zero/astronomical shape values that cause
    catastrophic floating-point cancellation.
  - Compute Burr log-density and log-survival using stable log1p-exp
    expressions.
  - Construct deterministic fitted severity-plot probabilities from
    conditional log-survival ratios rather than 1 - F(d) to avoid loss
    of tail precision.
  - Candidate fitting reports failed families as missing AIC/BIC rather
    than rewarding degenerate fits.
  - Add tests for heavy-censoring, large-scale covariates, nested
    intercept likelihood, Burr degeneracy, and extreme-tail plot
    survival.
  - These changes prioritize fit validity over the fastest possible
    fitting time. Source is distributed for independent R CMD check;
    CRAN checks still required on a machine with R.

                        Changes in version 0.2.1                        

  - Corrected roxygen2 help text so mathematical expressions are not
    interpreted as cross-reference links. Qualified utils::head/tail and
    made ggplot data-mask references explicit for clean R CMD check.

  - Corrected invalid reserved-word argument names in
    all-fixed-parameter fitting branches in fit_severity() and
    fit_spliced_severity().

  - Automatic numerical parscale for estimated scale intercepts;
    explicit optimizer overrides preserved, including in spliced models.

  - Candidate comparison now includes individual runtimes, optimizer
    evaluation counts and convergence codes.

  - Reuse repeated deductible/scaling term probabilities within each
    ordinary likelihood evaluation.

  - Candidate and single-fit plots default to deterministic
    censoring/truncation-aware bin/CDF summaries rather than simulating
    large portfolios. method = "simulation" retains the previous option.

  - fit_plot_data() and candidate_plot_data() support output =
    "plot_ready"; scaling_plot_data() exposes the exact long plot data.

  - User-facing plot axis-label sizes, angles and explicit tick
    breaks; 45-degree tick rotation for faceted candidates.

  - Exact rejection simulation for inverse Gaussian draws conditional on
    a deductible, avoiding most per-row inverse-CDF root searches.

  - New regression tests and documented benchmark script; updated
    README/vignette/examples.

                        Changes in version 0.2.0                        

  - Documentation-only CRAN check cleanup: unique S3 aliases, correct
    plotting helper documentation association, and wrapped candidate
    usage lines.

  - Optional deductible and censoring-limit columns for single,
    candidate and spliced fits.

  - One-pass anomaly screening, warnings, original-position
    excluded_rows, and reason table excluded_details; invalid_rows =
    "error" available.

  - Splice threshold search reports eligible body/exact-tail counts and
    explicit rejected-candidate reasons; configurable minimum uncensored
    tail observations.

  - Vectorized analytic limited expected values for exponential,
    Weibull, gamma and lognormal families; batched severity factor
    grids, plus paired scenario mode.

  - Log10 severity axes and xlim controls; candidate overlays/facets
    using observation-aware simulated claim data.

  - Expanded quick-start, worked candidate vignette, and API examples.

  - The package source is provided for local R checks. Its CRAN status
    is not certified in an environment without R.

                        Changes in version 0.1.0                        

  - Numerical warnings from invalid optimizer trial points are now
    handled internally; such points receive a penalized likelihood while
    invalid final fits still fail explicitly.

  - Initial release.

  - Fits common actuarial severity distributions to ground-up losses
    with deductible truncation and limit censoring.

  - Supports optional linear severity scaling, likelihood weights, fixed
    parameters, candidate comparison, and spliced body-tail models.

  - Generates limited expected severities, LERs, ILFs, deductible
    factors, combined rating factors, and valuation-adjusted expected
    payments.

  - Includes simulation-based goodness-of-fit diagnostics, scaling
    diagnostics, layer diagnostics, Hessian inference, and basic
    bootstrap support.

  - 0.2.1 source maintenance: resolved candidate plotting default
    argument handling within lapply() and consolidated roxygen2
    source-of-truth maintenance for help pages, NAMESPACE, package docs,
    and S3 print registration.