Module model 

The Boltzmann posterior over genomes, as a fugue program

Fix a prior program p(x) (a GenomePrior) and an observation program p(data | x) (a GenomeLikelihood — observe statements, latent nuisance parameters, and/or black-box factors). The target

    π_β(x) ∝ p(x) · p(data | x)^β

is literally a fugue model: prior.model().bind(|g| likelihood.model(&g, β).map(|_| g)). Every density this layer needs — prior mass, tempered joint, MH acceptance — is obtained by running or replaying that program; there is no hand-written density code.

For the classical black-box case, EvolutionModel::new(prior, fitness) wraps a scalar Fitness in FactorFitness (factor(β·f(x)) — the Gibbs / generalized-Bayes posterior); EvolutionModel::from_likelihood accepts any observation program.

Two builders exist because MH wants a fixed-β target while tempered SMC must receive the β = 1 program (fugue’s adaptive_smc supplies β by tempering log_likelihood + log_factors; baking β in as well would double-count it).

Structs

EvolutionModel

A probabilistic model of an evolutionary population: a genome prior program, an observation program (likelihood), and an inverse temperature β.