Genetic analyses:
We used CERVUS 3.0 (www.fieldgenetics.com) to calculate summary statistics (e.g., number of occurrences of each allele at each locus) and assess the suitability of loci for parentage analysis. CERVUS was then used to calculate the following four measures for each of the 10 high CO2 replicates: relative allelic richness (i.e., NAlleles, replicate/NAlleles, total), observed heterozygosity, number of dams, and number of sires. Linear regression was used to relate the four proxies to survival. Subsequently, CERVUS was used to assign individual offspring to candidate mothers and fathers. For each offspring tested, candidate parents were assigned with at least 95% confidence using exclusion probability first and the maximum likelihood score second. Comparing each offspring to all parent genotypes, exclusion of candidate parents followed if mismatches occurred at more than one locus. In total, 704 of 772 individuals (91%) were successfully assigned to parents, a success rate comparable to other studies on fish (75–95%).Unassigned offspring, resulting from poor amplification at some loci, were excluded from further analyses. All females, but only 35 of 42 males, were assigned to offspring, likely because the sperm from seven males was not activated or the fertilized eggs were not selected for in the experiment.

Last, we ran a hierarchical set of univariate ‘animal models’ for the trait DS, using the restricted maximum likelihood (REML) software ASReml v3.0.5 (VSN International Ltd, Wood Lane, Hemel Hempstead, UK). An ‘animal model’ is a mixed model (i.e., a form of linear regression with ‘fixed’ and ‘random’ effects as explanatory variables) that has been found advantageous for estimating phenotypic variance components in wild populations. For fixed effects, we only considered replicate ID (1–10), which corresponds to the experimental vessel, given that age (all individuals had the same day of fertilization) and sex (individuals sampled in their entirety) were not relevant factors. Replicate ID significantly influenced DS (df = 9, P < 0.001, univariate general linear model SPSS; IBM) therefore all subsequent analyses included replicate ID as a fixed effect to remove its effect prior to the estimation of genetic parameters. We then estimated the heritability of DS by testing three distinct models: (i) the base model including only the fixed effect, (ii) a model including the fixed effect and an additive genetic random effect (αi with variance VA), and (iii) a model including the fixed effect, additive genetic random effect, and dam identity (mi with variance VM) as a random effect. The components of total phenotypic variance in DS (VDS = VA + VM + VR) were then used to calculate narrow sense heritability h2 of DS as the ratio of additive genetic variance to total phenotypic variance (h2 = VA/VDS) and maternal effects m2 of DS as the ratio of maternal variance to total phenotypic variance (m2 = VM/VDS). Statistical significance was assessed with likelihood-ratio tests [LRT = −2 * (likelihood base model–likelihood model with random effects)] tested against the chi-square distribution (df = 1).

BCO-DMO Processing:
- Replaced spaces with underscores in species column.
- Replaced blanks with 'nd' to indicate 'no data'.
- Modified parameter names to conform with BCO-DMO naming conventions.
- Changed positive degrees west (long) to negative degrees east.