Evolutionary Population Databases: Populations as Dynamic Information Structures and Their Relation to Inverse Evolutionary Inference (IEI)

Evolutionary Population Databases (EPD) propose a new model of persistence inspired by population dynamics, in which durability and consistency are maintained at the distributional level through controlled turnover rather than by eliminating errors in individual records. Population distributions can retain information about the environments and histories that generated them. This paper introduces Evolutionary Population Databases (EPD), an engineered framework in which a population-valued state is treated first as a persistent data store and only secondarily as a site of computation. We separate a base layer, which preserves records by deterministic retention and neutral resampling, from a derived search layer, in which mutation, recombination, and selection operate on an isolated fork that never writes back. For the base layer we state a statistical consistency contract: a replication dial interpolating between an archival store and a stochastic store, with an exact expected pairwise-diversity recurrence and an exponential large-N_e approximation; probabilistic bounded-deviation reads; content-addressed deduplication with multiplicity; and immutable generation-snapshot semantics. A streaming instance adds an ingest rate μ: μ sets the recency horizon, whereas λ controls additional resampling noise and stochastic support compression. We derive the corrected stationary variance and a finite-sample heavy-hitter error bound. We then relate EPD to Inverse Evolutionary Inference (IEI). In an ensemble biallelic instance, the base layer stores population-level allele-frequency states; the exact stationary law point-identifies selection within a fixed one-parameter model, while the fitness optimum identifies only its sign. Finite stationary samples carry strictly positive Fisher information about selection magnitude, whereas the optimum carries none. The natural resultset of an evolutionary query is therefore a distribution rather than a single record, but the strength of the resulting inference is explicitly model- and observation-dependent.