                 Changes in version 1.6.3 (2026-09-20)                  

  - Move most vignettes to web_only vignettes directory, and update
    interface for visreg in mgcv.Rmd vignette

                 Changes in version 1.6.2 (2026-07-14)                  

  - Replace array with tmbutils::array in CPP, to disambiguate new array
    typedef in RcppEigen, following andrjohns in dsem#63
  - Remove "test-basic-fits.R" test "tinyVAST works as dsem" from CRAN,
    due to weird and un-reproducible ERROR on CRAN openBLAS check
  - Update epsilon_distribution to pass arguments using const and
    pointers (pass-by-const-reference) to avoid compiler errors
  - Move helper functions to tinyVAST namespace

                 Changes in version 1.6.1 (2026-06-30)                  

  - Fix AR sign in SAR example (was using negative, because I hadn't
    updated after changing signs in dsem)
  - Tag weights_i with DATA_UPDATE to explore SGD
  - Fix bug in predict when using sf_areal_domain input
  - Add error when rowSums(A_is)!=1 (i.e., samples are outside domain)
    and warning when rowSums(A_gs)!=1 (i.e., predictions are outside
    domain)
  - Add sem_imports.R to avoid listing sem in IMPORTS (allowed given
    that sem has GPL-3 licence), and based on prior email.

                 Changes in version 1.6.0 (2026-05-19)                  

  - Adding nearest-neighbors Gaussian Process option
  - Add tinyVASTcontrol(tmb_random) argument to allow exploration of
    penalized likelihood
  - Modifying tinyVASTcontrol(tmb_par) to allow users to supply a
    partial list of starting values, e.g., to provide values for
    semi-variance parameters when changing model resolution.
  - Add development$optimizer argument for using
    optim(method="L-BFGS-B") or TMB::newton, which is necessary when
    using penalized likelihood with >40,000 random effects. And removing
    tinyVASTcontrol option for suppress_optim_warnings (making it always
    TRUE)
  - Adding Makevar -DTMBAD_INDEX_TYPE=uint64_t to allow GetTape for RTMB
    manipulation
  - Change SAR using areal_model for spatial_domain to (1) eliminate
    row-standardization (which messed up the areal interpretation), and
    (2) defining an informative start for kappa to start with a value
    that avoids numerical under/overflow;
  - Adding argument block to integrate_output to allow calculating
    multiple derived quantities using a single call (to avoid extra
    MakeADFun calls)

                 Changes in version 1.5.1 (2026-04-14)                  

  - Fix bug, where previous versions where not simulating measurement
    errors for delta-models using obj$simulate(). h\t github ilar-cloud
    for reporting on issue-402

                 Changes in version 1.5.0 (2026-03-23)                  

  - Add starting value log_kappa when using SPDE method, using SD of X
    and Y coordinates and converting to equivalent kappa
  - Fix bug in project for logical class in empty vectors
  - Add development options to explore hess-compress option (and
    commenting out incomplete Cholesky for now)
  - Add sanity function to check for evidence of non-convergence
  - Improve detection of sf class for areal models

                 Changes in version 1.4.0 (2025-12-20)                  

  - Adding student distribution, borrowing/importing code from sdmTMB
  - Changing deviance_explained to map off student DF at MLE to avoid
    negative values, in way that can be easily generalized.
  - Changing deviance_explained to map off overdispersion for negbin1
    and negbin2
  - Changing deviance_explained to map of power parameter for tweedie
    distribution
  - add testthat checks comparing tweedie and negbin2 deviance-explained
    for tinyVAST with mgcv package
  - Add vignette for seasonal spatio-temporal index standardization,
    using Yellowtail flounder data from VAST demo

                 Changes in version 1.3.0 (2025-09-13)                  

  - Adding add_vertex_covariates to associate mesh vertices with
    covariates (h\t Sean Anderson)
  - Adding option for covariate-based anisotropy using vertex_formula,
    while also allowing geometric anisotropy.
  - Adding option for barriers using triangle_formula which permits an
    offset for barrier covariates
  - Adding spatial_cor to use sparse matrices to compute the correlation
    between a coord and pred coordinates, e.g. to visualize
    covariate-based anisotropy
  - Fix make_eof_ram to work as intended for multivariate models, e.g.,
    having a separate response map and shared indices across variables
    (previously had a shared response map and shared indices, i.e.,
    collapsed to univariate model)
  - Fix bug arising in predict when spatial_varying involved a column of
    newdata that was expecting a factor, but provided a
    character-vector. Now coercers to factors using the levels of the
    original data (same behavior as for formula)
  - Export fit$rep$negloglik_i as log-likelihood for each datum, for use
    in calculating out-of-sample predictive score
  - Turn off SE reporting during index calculations (as speedup), and
    allow SE reporting for mu_g in predict
  - Fix bug arising in project when spatial_varying was specified
  - Adding web-only vignette showing predator-expanded diet estimator
    using joint model of specific stomach contents and predator density
  - Fix bug arising when using weights argument using family of
    delta_gamma or delta_lognormal, where this bug was introduced in
    release 1.2.0 (h/t Peri Gerson for identifying the issue)
  - Add vignette showing project function
  - Add a web-only vignette showing the sponge-coral-fish case study to
    illustrate spatial structural modelling

                 Changes in version 1.2.0 (2025-07-19)                  

  - Adding a new spatial domain using sf::st_make_grid
  - Adding option for geometric anisotropy when using sf::st_make_grid
  - Fix bug in deviance for lognormal distribution in delta models (h/t
    Sean Anderson for finding it)
  - Change family = "binomial" to use weights as number of Binomial
    trials N
  - Fix deviance_explained to work with non-default weights argument
  - Add fit$internal$packageVersion to allow predict, cAIC etc to check
    and throw error if there's a package inconsistency between object
    and installed package
  - Add conditional_gmrf to do conditional simulations from a GMRF
  - Add project to project tinyVAST forward in time
  - Eliminate ... argument to tinyVAST so that it's obvious if an
    argument is mis-named.
  - Fix bug which previously resulted in non-converged models, where
    including DSEM arrow "x<->x, 0, sd_xy" would improperly add another
    term "x <-> x, 0, V[x]" because it didn't previously parse the
    whitespaces when detecting missing covariances (h/t Jon Reum for
    flagging the issue)

                 Changes in version 1.1.1 (2025-05-07)                  

  - Adding option for geometric anisotropy when using the SPDE method
  - For some reason, the CPP edits also address an error message during
    devtools::check_win_devel "array subscript 'const __m128i[0]' is
    partly outside array bounds of 'unsigned char [12]'"

                 Changes in version 1.1.0 (2025-05-03)                  

  - Adding term_covariance to calculate the covariance among variables
    for SEM term, or covariance among variables-and-lags for DSEM terms
  - Adding Restricted Spatial Regression estimator for covariates
    alphaprime_j and alphaprime2_j to fit$rep output.
  - Adding methods to allow use of cv to calculate crossvalidation skill
  - Add bias.correct option to predict (but still no flag for SEs for
    anything except p_i)
  - Switch sample_variable from using obj$env$MC to
    obj$env$spHess(random=TRUE) which seems more stable as dependency
  - Add functionality for te and ti splines, although they remain poorly
    tested
  - Add error check to sfnetwork_mesh to detect if the stream network is
    not ordered as a tree
  - Improve stream network vignette to use matrix notation for joint
    precision, and modify simulate_sfnetwork to use that
  - Change tinyVAST.cpp to use matrix notation constructor and fix bug
    in previous constructor where the covariance between first and
    second nodes was not right
  - Expand test-sfnetworks.R integrated test to confirm that
    matrix-notation precision constructor is identical to the inverse of
    Ornstein-Uhlenbeck covariance as intended.

                 Changes in version 1.0.1 (2025-03-21)                  

  - Modify examples for simulate.tinyVAST and sample_variable to try to
    avoid terminal output giving error in valgrind check
  - Add ivector_minus_one function to satisfy clang-UBSAN
  - Swap GMRF(Q).Quadform(x) to x.matrix().transpose() * (Q *
    x.matrix()) to avoid calculating log-determinant of Q in the
    smoothers penalty, to avoid valgrind errors
  - Add tinyVASTcontrol option that suppresses nlminb warning messages
    by default, which are typically not informative to casual users

                 Changes in version 1.0.0 (2025-03-13)                  

  - Adding nbinom1 and nbinom2 families
  - Simplify argument names, by changing sem to space_term, dsem to
    spacetime_term and spatial_graph to spatial_domain, and eliminating
    delta_ in the names for arguments to delta_options
  - Add time_term to allow time-variable interaction (e.g., AR1
    intercepts)
  - Adding overview and model-description vignettes
  - Add simulate S3 generic

                        Changes in version 0.7.1                        

  - Fixed bug (wrong output) when using predict(fit, what="mu_g") and a
    Poisson-linked delta model
  - Fixed bug (cryptic error message) when using integrate_output
  - Add cAIC (but disabling EDF calculation for now)

                        Changes in version 0.7.0                        

  - Adding option for spatially-varying-coefficient (SVC) models
  - Add error-check for when data has a factor with extra levels, which
    conflicted with the logic of adding all origdata levels to newdata
    when calling predict, and hence caused an uniformative error
    previously

                        Changes in version 0.6.0                        

  - Change integrate_output interface by splitting W_gz and V_gz into
    four vectors area, type, covariate, and weighting_index to simplify
    documentations and improve naming
  - Fix bug where cloglog and logit links were not previously
    implemented for use in predict and integrate_output

                        Changes in version 0.5.0                        

  - Adding vignette showing how to fit multiple data types in an SDM
  - Adding deviance_explained and calculating this by default

                        Changes in version 0.4.0                        

  - Adding code for simulation residuals, and examples to vignettes

                        Changes in version 0.3.0                        

  - Adding sdmTMB as dependency, and importing family options from it
  - Adding vignette for joint analysis of condition and density

                        Changes in version 0.2.0                        

  - Add option to specify covariance in SEM and DSEM notation

                        Changes in version 0.1.0                        

  - Initial public alpha release
