Incorporates population-specific, seasonal, and discarding dynamics.
Major changes
Population structure and dynamics
Added population-specific (natal homing) and seasonal dynamics to both simulation and estimation. Most dimensions now follow population, region, year, season, age, sex, and fleet.
Added movement-after-mortality and continuous movement dynamics.
Added continuous-time Markov chain (CTMC) movement estimation using preference functions.
Added move_expm_nsub to Setup_Mod_Movement, allowing CTMC matrix exponentials to be evaluated using implicit backward Euler substeps, providing a faster alternative for differentiation while preserving non-negative fractions and exact survivor catch balance.
Added a state-space treatment of the numbers at age via NAA_re in Setup_Mod_Biologicals and Setup_Sim_NAA_state. The log numbers themselves become random effects for ages two and older including the plus group, with the deterministic mortality and ageing step as the prediction they are evaluated against. Forms are "iid", "1dar1_a", "1dar1_y", "2dar1", "3dcond" and "3dmarg" over the age-year grid, composable with unstructured correlations across populations, regions, seasons and sexes.
The state runs at season boundaries as well as year boundaries. NAA_re_seasons takes "annual" (the default, a state at season one only, so the numbers within a year stay deterministic and the innovation is purely annual), "all", or an arbitrary set of season indices, which need not be contiguous so that states can be placed only in the seasons with observations. NAA_re_season adds an unstructured correlation across the active seasons, and NAA_sigma_seasblk_spec blocks the process error standard deviation over them. ln_NAA, ln_sigmaNAA, naa_sigma_blocks, NAA_pred and NAA_scalar all have the season dim between year and age; saved input lists without it are promoted and hold the state at season one.
The state-space numbers at age are not projected. Do_Population_Projection() advances projected numbers deterministically from the terminal year with recruitment the only stochastic element, so a forecast from a state-space fit omits that process error. The closed loop operating model does advance the state forward.
Fixed the tag cohort rescale under a state-space fit: it indexed the state factor by years at liberty rather than by calendar year, so it was correct only for cohorts released in the first model year.
Added support for local density dependence in reference points for meta-population and natal homing models.
Added uncertainty estimation for reference points.
Data and observation processes
Added population-specific catches, indices, compositions, tagging data, discards, and discarded compositions to simulation and estimation.
Recoded tagging to allow fleet-specific reporting rates and missing attributes in tagged fish. conv_tag_t_tagging, ln_init_conv_tag_mort, and ln_conv_tag_shed are now per-release-event vectors rather than global scalars; scalars remain supported and are recycled. init_conv_tag_mort_spec/conv_tag_shed_spec accept "fix", "est_shared", or "est_all".
Added internal OSA residuals for catch, indices, compositions, and tagging data.
Added support for at-age data sources (e.g., ICES-style stock assessments).
Recruitment and fishing mortality
Added age-0 (rec_lag = 0) Beverton-Holt recruitment, driven by the same year’s SSB.
Added sr_penalty ("none", "bh", "ricker") to fit a stock-recruit relationship as a penalty rather than as the recruitment process itself. sr_R0_spec controls the scale of the stock-recruit curve. MSY reference points remain unavailable under mean recruitment.
Added Fdev_model ("iid", "rw", "ar1") to Setup_Mod_Catch_and_F, with Fdev_rho/Fdev_rho_spec for the AR1 case.
Fishing mortality deviations now distinguish missing catch observations from true recorded zeros: NA in ObsCatch with UseCatch == 0 estimates F as an active year, whereas an observed zero forces Fmort to zero.
Added the ability to parameterize fishing mortality as free annual log-F via ln_F_mean_spec = "fix".
Biological processes and growth
Growth parameters can vary over time via growth_tv_model ("iid"/"rw"), with support for year- and cohort-specific variation.
Added semi-parametric growth (growth_semipar: "iid", "rw", "2dar1", "3dmarg", "3dcond") for year-by-age deviations in mean length at age.
Added the Richards growth form (growth_model = "richards"), with von Bertalanffy growth recovered when rho = 1.
Added flexible size-age and weight-at-age inputs, including fleet-specific size-age keys and selection-weighted weight at age.
Added conditional age-at-length fitting and Francis reweighting.
Selectivity
Added non-parametric, bicubic, logistic-with-asymptote, and fixed selectivity forms.
Added sex-specific selectivity offsets, allowing selectivity parameters to be linked across sexes using "par", "scale", "par_scale", "apical", and "par_apical" forms.
Added time-, bin-, and surface-based selectivity deviations and expanded selectivity smoothness penalties.
Corrected the double-normal ("dbnrml") selectivity parameterization and added flexible anchoring of its ascending limb.
Added _NSelBins_<n> suffixes to selectivity model strings to hold bins beyond n at the value of bin n.
Added fish_sel_norm_bins/srv_sel_norm_bins to specify the bins over which non-parametric log-scale selectivity is standardized.
Setup_Mod_Rec has a new R0_blocks argument, which takes the same string forms as the selectivity blocks ("none_Pop_p", "Block_b_Year_a-e_Pop_p") and gives ln_global_R0 a block dimension, together with R0_ref_block for the single value the initial age structure, the R0 prior, the ln_rinit penalty and the stock-recruit scale all read. Under mean recruitment an R0 block is a productivity regime; under a stock-recruit form it is the curve’s scale. R0_yr is reported. A model without R0 blocks keeps one column and is byte-identical to before.
An analytically solved catchability is now solved within each catchability time block rather than pooled over the whole series. srv_q_type / fish_q_type of "arith" or "geo" previously computed one value per region and fleet and discarded srv_q_blocks / fish_q_blocks silently, so a blocked analytic q looked configured and did nothing. Each block is now solved from the observations it owns, a block with none takes the pooled value, and a fleet with a single block is unchanged to machine precision.
Reference points
Reference points now support local density dependence for meta-population and natal homing models.
Added uncertainty estimation for reference points.
Get_Reference_Points now uses growth-derived weight at age from the model report rather than the data placeholder.
Get_Reference_Points() has a new calc_rec_yrs argument, the model years mean recruitment is averaged over for SPR biomass reference points, as an alternative to the calc_rec_st_yr:(n_years - rec_age) window, so the mean can run to the terminal year or skip years. It is passed through get_key_quants(), get_closed_loop_reference_points() and Get_Reference_Point_Uncertainty(). The mean recruitment is also now returned as mean_rec[n_pop, n_regions] under every method.
Case studies and applications
Added BSAI Atka mackerel, West Coast sablefish, EBS Pacific cod, North Sea sand eel, GOA Rex Sole, BSAI Northern Rock Sole case studies bridging recent assessments.
Model fitting and diagnostics
Added dynamic structural equation models via Setup_Mod_DSEM, linking recruitment, growth, movement, catchability and the numbers at age to each other and to covariates through arrow notation.
Setup_Mod_Movement and Setup_Mod_Tagging no longer need to be specified; fit_model fills their off settings, movement for one region only.
Added bias_correct_pe and bias_correct_oe to Setup_Mod_Dim and Setup_Sim_Dim. bias_correct_pe takes "none", "rec" (the default, and what every model did before it existed) or "all", and decides which process deviations are centered on minus half their marginal variance: recruitment and the initial ages under "rec", those plus the numbers at age state under "all". do_rec_bias_ramp and bias_year still scale the recruitment part. bias_correct_oe predicts a lognormally fit observation at its mean rather than its median, covering indices, catch, discards and the age-disaggregated sources.
Recruitment deviations under RecDevs_model = "ar1" are now mean unbiased, centered on minus half the stationary variance rather than on zero. A random walk takes no correction, since it has no stationary variance.
Setup_Sim_Rec gained sigmaR_switch, so the operating model reads the early and late recruitment sigma by year as the estimation model does. It previously used the late value in every year.
Improved Newton refinement in fit_model by obtaining the Hessian directly from the AD tape rather than finite differencing.
Added newton_hessian to fit_model. "exact" (the default) is the tape Hessian above. "diff" differences the exact gradient for models whose tape Hessian is too large to build, and in place of the Newton step reruns nlminb with each parameter in its own standard errors, which converges where a raw Newton step on a differenced Hessian goes astray. The last Hessian is returned as $hessian for sdreport.
Added ctmc_diffusion_bounds = "sg", the Scharfetter-Gummel flux with : diffusion and taxis go into one rate that is positive and smooth everywhere, central at small taxis and upwind at large. It has none of the softplus’s hinge, where an edge that taxis nearly cancels sits at a bend of width ctmc_diffusion_eps that the optimiser crawls along, nor its floor of eps * log(2).
fit_model no longer takes a Newton step that raises the largest absolute gradient. With an indefinite Hessian the step could leave the nlminb optimum for a worse point; refinement now stops there instead.
Added label_dims(), which returns a copy of the report, the parameter list or the data list with every array’s dims named (Pop, Region, Year, Seas, Age, Len, Sex, Fleet, and Block, Sel_Par, Bin and so on for block and parameter dims) and labeled with calendar years, model ages and length bins, so reshape2::melt() gives named columns. The lists themselves are left as they were, since the model and the plotting and diagnostic functions read their arrays by position.
Added 95% confidence intervals for SDNR based on a chi-squared test for OSA residuals.
Added OSA residuals for time-series observations, compositions, and tagging data.
Added data$extra_nll, a likelihood from outside the package: a function (pars, env) whose value is added to the joint likelihood (reported as extra_nLL), where env is the objective’s environment, so the function can read anything the model computes (movement, mortality, numbers at age, spawning biomass, selectivities, …).
Improvements
Numeric and character codes are both accepted for simulator dynamics, matching the estimation model.
Added a shared process-error sharing-spec helper (build_pe_map/build_shared_spec_map); migrated sigmaF_spec, sigmaC_spec, sigmaR_spec, Fdev_rho_spec, and do_Fmort_mapping onto it (for developers).
Refactored movement’s continuous process-error map and log-likelihood onto the same shared code (for developers).
Newton refinement in fit_model now takes its Hessian from the AD tape (obj$he) instead of finite-differencing, and stops early on a non-finite Hessian.
Index ranges built from length(), nrow(), ncol() and dim() now use seq_along() and seq_len(), so an empty object gives no iterations instead of the indices 1 and 0. Semicolon chains are split one statement per line, T/F are written as TRUE/FALSE, two if() conditions use &&, operators have spaces around them, blocks that sat 2 or 4 spaces off their braces are re-indented, and multi-line pipelines put each step on its own line (for developers).
catch_to_F_singlefleet() and catch_to_F_multifleet() are replaced by catch_to_F_om(), which converts catch advice by region, season and fleet into the operating model’s F by running the year through the operating model’s own dynamics, solved with Newton steps on an RTMB::MakeTape() tape of each season’s catch.
Bug Fixes
Fixed the time-varying selectivity smoothing penalty not being applied in the first year.
Fixed the projection dropping the AD class when joining assessment and projected spawning biomass for the stock-recruit curve, which took recruitment off the tape.
Fixed do_runs_test() returning a NaN p-value when the runs test has too few runs for a variance (a two-point series, or zeros with every other residual one sign). It now returns 0.001, as it already did for a series with no sign changes.
Fixed the aggregated Dirichlet-multinomial composition likelihood (_Type = 0) not adding addtocomp to the expected and observed proportions.
version 1.1.0
Release Date: 2026-3-31
Major changes
Added capability to estimate movement using continuous time Markov chains (CTMC) using preference functions.
Bug Fixes
Fixed bug to allow for size-based selectivity in the estimation model.
version 1.0.0
Release Date: 2025-11-24
Major changes
First stable public release of SPoRC.
Introduced a modular framework for estimating age-, sex-, season-, and region-structured population dynamics.
Added closed-loop simulation and management strategy evaluation (MSE) capabilities.