Sea-temperature effects on recruitment of 282 RAM Legacy stocks rarely replicated, as weak constant effects predict, and gave no forecast gain; later effects diverged toward losses at warm range edges

Question: do sea-surface temperature (SST) effects on fish and invertebrate recruitment that are fitted in an early period hold later and improve forecasts? Myers (1998) found that most environment-recruitment correlations failed when tested again, except at range limits. Data: RAM Legacy v4.66 recruits and spawning biomass, HadISST1 and ERSST v5 SST in hand-made stock-area boxes, and OBIS occurrence grids for each species' thermal range. Method: a Ricker model with brood-year SST per stock and period, in three splits: series halves (282 stocks, 142 species, brood years 1900-2022), before and after 1995 (217 stocks), and a real-time split that fits older RAM assessments and tests only newer brood years of the newest assessment (153 stocks). We compared observed statistics with stationary simulations (constant effects, AR(1) errors) through the identical pipeline. Results: in the half split, 35 stocks (12%) had an early effect at p < 0.05 and 6 of these (17%) replicated. Simulations with shrunk (realistic) true effects expected 31 (22-41) and 17% (7-28%); simulations that took each stock's full-period effect as true expected 57 and 34%. Thus low replication alone is what weak constant effects give. But the change between periods was larger than stationary: mean z-squared of the change was 1.72 (expected 1.29, 1.08-1.51) and the between-stock spread of true late effects was 0.25 per C (expected 0.17, 0.13-0.22); the calendar split agreed. The late spread followed thermal position: warm-edge stocks had more negative SST effects (slope -0.30 per C per unit position, SE 0.10; early -0.04, SE 0.06; within-region permutation p = 0.03), but region fixed effects reduced it by 30% (p = 0.12). Myers' range-edge exception was not supported (20 edge stocks, no replication). Stock-specific SST terms made late forecasts worse (mean log MSE ratio +0.042, 95% CI 0.013-0.072; +0.06 in the real-time split); shrunk terms and a thermal-position rule were neutral, and in 93% of stocks a correct SST term could give less than 1% gain. Limitations: assessment-model recruitment, approximate boxes and annual surface SST, survey-biased OBIS ranges, few edge stocks, and exploratory change diagnostics added after the first results.