Who Is Missing from Your Simulated Society? Toward Omission-Aware Population Validity in LLM Social Simulation

LLM-driven social simulations are increasingly used to study opinion dynamics, test platform interventions, and pilot policy. The validity debate around them has centered on fidelity: whether simulated personas reproduce the attitudes and behaviors of the humans they represent. We argue that fidelity evaluation is conditional on inclusion—it can assess whether represented agents behave accurately, but it cannot by itself establish whether the simulated population supports claims about the target society. We distinguish coverage omission from underrepresentation and representation error; trace an omission cascade from target population through sampling frame, persona base, generation, and interaction to the aggregate claim; and give a tipping-point analysis showing that plausible assumptions about excluded populations can reverse the sign of a simulated intervention effect at realistic exclusion rates. Because validation data typically shares the selection mechanism of construction data, fidelity checks cannot detect these failures on their own. We propose an omission-aware population-validity protocol: population-and-claim statements, independent absence audits, sensitivity bounds over plausible excluded populations, and population-level omission-awareness reporting. Unlike demographic-representativeness reviews, validity principles for LLM societies, and simulation-boundary analyses—which examine within-frame fidelity and heterogeneity—our target is population-frame validity. The goal is not complete simulated societies, which are impossible, but simulated societies whose incompleteness is declared, bounded, and priced into every downstream claim.