In models of natural selection, fitness is often defined as an organism’s number of descendants. Due to stochastic factors, this number is commonly treated as a random variable. Proponents of the Propensity Interpretation of Fitness (PIF) acknowledge this by defining fitness as a token organism’s propensity to leave descendants. However, as Sober (2022) has argued, such single-case propensities are epistemically inaccessible: we cannot reliably estimate them when each organism provides only one “trial.” While some authors suggest that cloning experiments might help secure the epistemic accessibility of token fitnesses, Eagle (2004) has shown that experimentally generated frequencies alone cannot pin down single-case propensities, as different propensity assignments can be compatible with the same frequencies.

In response, this talk proposes a “long-run” propensity interpretation of fitness, defining fitness as the propensity of organism types to produce a certain number of descendants in the long run, with a specific limiting frequency. With suitable inductive uniformity assumptions, this approach allows empirical testing of fitness-value hypotheses through their inductive consequences for finite frequencies. By showing how to apply fitness measures so construed to prima facie problematic cases such as cyclic or chaotic evolutionary dynamics, I defend this approach against various objections in the literature.