Paper4 (exp-evo-loop-pilot) showed DECLINING novelty trajectory when LLM is the substrate. This pilot tests the complementary architecture: Conway's Life as substrate, LLM agents as evaluator panel, evolutionary selection on CA seeds. CA has genuine computational irreducibility — novelty source is outside LLM training distribution. LLM evaluates but does not generate. Falsifiable prediction: if LLM interestingness scores CLIMB across generations (agent consensus improves), external-substrate + LLM-evaluation unlocks novelty that LLM-as-substrate cannot. If scores plateau/decline: LLM evaluator itself is retrieval-bounded (evaluates against same training templates regardless of CA).