Expand description
The learned taste as a fugue-evo fitness: the surrogate that lets the machine evolve thousands of candidates silently (the reference: The two loops).
SurrogateFitness plugs the taste posterior’s expected utility into
fugue-evo’s Fitness, so the Boltzmann target
π_β(x) ∝ p_grammar(x) · exp(β · E[u_θ(φ(x))]) becomes an ordinary
EvolutionModel and typed-MH / SMC drivers apply unchanged. Quarantined
candidates score a large negative fitness — safety layer 2: evolution
learns to avoid the pathological region.
Structs§
- Surrogate
Fitness - Expected posterior utility as a scalar fitness over patch terms.
Constants§
- QUARANTINE_
FITNESS - Fitness a quarantined (unrenderable/unlistenable) candidate receives.