EBS Pollock RTMB Bridge

The exact September 2025 numerical reference and its regression gate are documented in BRIDGE_PROVENANCE.md. Post-bridge changes and the first identified numerical regressions are documented in POST_BRIDGE_CHANGE_AUDIT.md.

The primary rendered report is built from reporting/sept_2025_bridge_restoration.qmd and includes lightbox-enabled figures and a render-time execution of the bridge gate.

This repository contains a standalone RTMB reimplementation of the ADMB EBS pollock bridge model. It is intended for model-port debugging, reproducibility checks, and diagnostics comparing RTMB output to a dedicated ADMB bridge run.

The repository contains source code, reporting scripts, and the minimal ADMB bridge bundle needed for standard RTMB runs. Generated RDS files, rendered HTML reports, full ADMB run products, and projection executables are intentionally not versioned here.

Bridge Inputs

The default run uses the bundled bridge files in this repository:

• admb/runs/for_rtmb/pm.par
• admb/runs/for_rtmb/pm.rep
• admb/runs/for_rtmb/pm.tpl
• admb/runs/for_rtmb/pm.dat
• admb/runs/data/

To compare against a different ADMB bridge run, point the repository at another pollock workspace with:

export POLLOCK_ROOT=/path/to/pollock

For compatibility with older scripts, POLLOCK_BASE is also recognized when POLLOCK_ROOT is not set. If neither variable is set, the code uses the in-repository admb/runs/ bundle.

Run

From this repository root:

Rscript R/write_output.R

The standard output is written to:

analysis/output/base.rds

To fit the model directly:

Rscript analysis/Run_rpm.R

To render the diagnostics report after generating required outputs:

quarto render reporting/ebs_pollock_rtmb_ebswp_assessment.qmd

To rebuild the fixed-parameter September 2025 bridge and every published downstream product from the same base file—including the bridge test, nine-peel retrospective, both projection suites, tuned SparseNUTS run, and Quarto report—run:

Rscript R/rebuild_bridge_products.R

The rebuild stops if any downstream product records a different base.rds checksum.

Likelihood profiles

The profile workflow fixes a selected parameter at each grid point and either reoptimizes all remaining parameters (a likelihood profile) or evaluates the objective without reoptimization (a slice). It records the total objective, optimizer diagnostics, and the contribution from each reported likelihood component.

Run the default 17-point reoptimized profile for log_avgrec with:

Rscript R/run_likelihood_profiles.R

Configure a run with environment variables. Repeated RTMB parameter names must use one-based occurrence notation such as log_rec_devs[10].

PROFILE_PARAMETERS="log_Rzero,log_q_ats" \
PROFILE_POINTS=11 \
PROFILE_HALF_WIDTH=0.3 \
PROFILE_MODE=reopt \
Rscript R/run_likelihood_profiles.R

When a requested parameter is fixed in R/config.R, the profile runner rebuilds the RTMB map with that parameter released before fitting the profile base. This is the default behavior for log_avgrec; it changes the diagnostic configuration but does not alter the standard bridge configuration used by other scripts.

The log_avgrec default spans plus or minus 2 units on the log scale. This wider range is intentional: the recruitment deviations compensate for modest changes in average recruitment, and the narrower plus-or-minus-0.35 trial reached only about 0.08 Delta NLL. The wider profile crosses 1.92 on both sides.

PROFILE_HALF_WIDTH is measured on the fitted parameter scale, which is the log scale for parameters whose names begin with log_. Outputs are written under the ignored analysis/output/profiles/ directory as RDS, CSV, and PNG files. The CSV records convergence codes and maximum absolute gradients; inspect these before interpreting profile shape. PROFILE_MODE=slice is useful for a quick code check but is not a replacement for reoptimization.

Profile figures use ggthemes::theme_few(), include the total objective, and use a common 0--2.1 Delta NLL scale for every facet. The Objective panel marks Delta NLL = 1.92, the usual approximate 95% likelihood interval threshold for one profiled parameter. Choose a grid wide enough to cross that threshold on both sides of the minimum; increase PROFILE_HALF_WIDTH when it does not.

The runner requires a base-fit convergence code of zero and a maximum absolute gradient no larger than 0.002. Increase PROFILE_MAX_EVAL (default 5000) if the optimizer stops early. PROFILE_GRADIENT_TOL changes that threshold. PROFILE_ALLOW_NONCONVERGED=true bypasses the check for exploratory debugging, but output from such a run should not be used as assessment evidence.

Time-varying double-logistic fishery selectivity

Form 2 is implemented as an optional two-stage model while the default September 2025 bridge path remains unchanged. Stage 1 estimates one double-logistic parameter triplet shared by all years. Stage 2 treats annual deviations from that triplet as random effects. The current tested treatment uses a 20% process CV and does not force ages 11 and older to share one selectivity value.

Run the focused implementation test with:

Rscript tests/test_double_logistic_form2.R

The complete experiment ladder is in R/run_double_logistic_experiments.R. Its outputs are written to the ignored analysis/output/double_logistic_experiments/ directory. The fitted comparison and selectivity surfaces are documented in docs/fishery_selectivity.html#fishery-selectivity-surfaces.

Run the focused helper tests with:

Rscript tests/test_profile_components.R

See REPRODUCIBILITY.md for the full workflow and expected software dependencies.

The rendered report is published with GitHub Pages at:

https://jimianelli.github.io/rtmb_ebswp/

Repository Contents

• R/: RTMB model, data utilities, run scripts, and output writers.
• analysis/: top-level analysis entry points.
• admb/runs/: minimal ADMB bridge inputs and comparison outputs.
• reporting/: Quarto diagnostic report source and run notes.
• data/README.md: description of bundled bridge data and override behavior.

Generated outputs are ignored by git and should remain under analysis/output/ or output/.

Status

This is a bridge implementation, not a final stock assessment product. Results should not be interpreted as accepted assessment advice without the usual convergence, estimability, retrospective, sensitivity, and review diagnostics.

License

Code in this repository is available under the MIT License. See LICENSE.md. External assessment data, ADMB bridge outputs, and generated products are not included in this license unless explicitly stated by their source.