<p>by Aleksei Belyi, Yiyang Du, Alastair J. Wilson, Ben Longdon, Francis M. Jiggins</p> Host–parasite coevolution is expected to generate strong selection for resistance and infectivity that has the potential to erode genetic variation. Despite this, natural populations often retain extensive genetic variation in these traits. Genotype-by-genotype interactions on fitness can generate negative frequency-dependent selection, which helps maintain this variation by preventing any host or pathogen genotype from retaining a permanent advantage. Because these interactions place genetic variance in the interaction term rather than in additive differences among hosts or pathogens, that variance is masked from directional selection and can remain ‘cryptic’. Here we combine coevolutionary modelling with large‑scale experimental infection assays to quantify how host-pathogen interactions structure fitness variance in a vertically transmitted virus system, <i>Drosophila melanogaster</i> and sigma virus. Simulations show that coevolution typically erodes additive genetic variance in host and pathogen fitness, concentrating variance in host-virus interaction terms. Consistent with these predictions, experiments spanning 90 host-virus genotype combinations reveal that transmission, viral load, and virulence are overwhelmingly governed by host-virus genetic interactions rather than host or virus main effects. As a result, neither host resistance alleles nor viral genotypes confer consistently higher fitness across genetic backgrounds. Interactions tend to be sex‑specific, further limiting heritable fitness variation. Our results demonstrate that coevolution can substantially mask heritable genetic variation from directional selection by rendering fitness context-dependent. This extensive cryptic genetic variation may be revealed either when ecological or evolutionary conditions shift, or when the process of coevolution itself alters the direction of selection. This demonstrates the need to account for genotype-specific interactions when forecasting evolutionary responses.