gamma1 (depletion)
gamma2 (escapement)
Selected gamma
Precautionary yield
Recruitment source
Summary statistics
Natural mortality M is not an input: it is derived per iteration by fitting the proportional-recruitment model. Equivalent to Table 2 of WG-FSA-2022 Appendix G.
Spawning stock status
SSB relative to the median unfished SSB0, unfished (gamma = 0) versus the selected gamma. Median with 90% (shaded) and 95% (dashed) intervals. Equivalent to Figure 1.
Spawning stock status at gamma1
Spawning stock status at gamma2
Individual trajectories
One row per run x year, as in Proj_output_48.1_*.rds.
GRYM splits cleanly at the maturity ogive: adults are the spawning stock (SSN), juveniles are the rest (N - SSN). Overlay observed juvenile/adult abundance from Ryabov et al. 2023 (PAL-LTER). Absolute scales are not comparable -- GRYM numbers are per-recruit -- so both are normalised to their own series mean.
How GRYM models recruitment, and what else it could be
This is the parameter catches are most sensitive to, so it is worth being explicit about what GRYM assumes. GRYM recruitment is density-INDEPENDENT: each year draws from a fitted distribution regardless of how many adults there are, until SSB falls below the depletion threshold, at which point it is scaled down linearly. It has no stock-recruit curve at all.
Stock-recruit relationship
Realised recruitment distribution
Recruitment variability through the projection
Four checks on the recruitment model, from 'does it fit what it was fitted to' through to 'does it predict years it never saw'. All four run off the current recruitment vector, so press Run projection first.
The fit targets one observed pair: the mean and variance of the proportion of recruits in survey samples. Simulate from each fitted (M, CV) and the cloud should land on that target. If it misses, the recruitment model cannot reproduce the statistic it was fitted to and nothing downstream is trustworthy.
prFit solves for two unknowns (natural mortality M and recruitment CV) from two numbers. Many combinations fit the same observed pair about equally well, so the pairs trace a ridge rather than a cluster. That matters here because M is DERIVED, not assumed: a wide ridge means the data barely pin it down. Pakhomov (1995) 0.5-1.1 shaded for reference.
GRYM draws each year's recruitment independently. Real krill recruitment is episodic and environmentally forced, so it is autocorrelated. No slider on this panel can reach this: it is structural. Compare the model's autocorrelation against an observed series.
The honest test. Fit the recruitment model to the EARLY part of an observed series only, then ask what it predicts for years it never saw. The proportion of recruits is taken as juveniles / (juveniles + adults) from Ryabov et al. 2023. If the held-out years land inside the predicted cloud, the model generalises; if not, it is fitting the training window.
One-at-a-time sweep: shift each parameter down and up by a fixed percentage, re-run the whole assessment, and rank the parameters by how far the selected gamma moves. This DEMONSTRATES which parameter the harvest rate actually depends on, instead of asserting it.
Two things to hold in mind before reading the ranking. (1) This ranks sensitivity of gamma, which is a rate. The catch limit is gamma x B0, and B0 comes from the survey (Module 1), so a parameter can look modest here and still dominate the tonnage. Recruitment's reputation rests partly on the biomass side, which this sweep holds fixed. (2) One-at-a-time walks the axes of the parameter space, not its interior, so two parameters that only matter jointly will both look unimportant. Raise the gamma grid points until the top swing clears the resolution floor reported below, or the ranking is just rounding.
Current values against the Table 1 baseline. Changed rows are flagged with their percentage shift.
Save the projection ensemble or the parameter set.
Ensemble (.rds) Ensemble (.csv) Parameters (.csv)