Inferring the genetic architecture of evolution in the fossil record is difficult because genetic crosses are impossible, the acquisition of DNA is usually impossible and phenotype–genotype maps are rarely obvious. However, such inference is valuable because it reveals the genetic basis of microevolutionary change across many more generations than is possible in studies of extant taxa, thereby integrating microevolutionary process and macroevolutionary pattern. Here, we infer the genetic basis of pelvic skeleton reduction in Gasterosteus doryssus, a Miocene stickleback fish from a finely resolved stratigraphic sequence that spans nearly 17,000 years. Reduction in pelvic score, a categorical measure of pelvic structure, resulted primarily from reciprocal frequency changes of two discrete phenotypic classes. Pelvic vestiges also showed left-side larger asymmetry. These patterns implicate Pitx1, a large-effect gene whose deletion generates left-side larger asymmetry of pelvic vestiges in extant, closely related Gasterosteus aculeatus. In contrast, reductions in the length of the pelvic girdle and pelvic spines resulted from directional shifts of unimodal, continuous trait distributions, suggesting an additional suite of genes with minor, additive pelvic effects, again like G. aculeatus. Similar genetic architectures explain shared but phyletically independent patterns across 10 million years of stickleback evolution.