Position: Multi-Agent LLM Simulation as Approximate Posterior Inference Demands a Probabilistic Calibration Standard

Multi-agent simulations driven by large language models (LLMs) are increasingly used to study emergent collective behavior. The standard reporting convention treats the empirical distribution of agent strategies as a direct measurement. We argue that this convention obscures a probabilistic structure: an LLM-driven multi-agent simulation is best understood as an approximate sampling procedure that produces draws from p(theta | E, c), the posterior over latent agent-population parameters theta conditional on environment E and prompt configuration c. The standards of probabilistic inference therefore apply: posterior contraction, prior sensitivity, simulation-based calibration (SBC), and identifiability. We support the position with three micro-experiments on a stylized 25-agent population: posterior contraction with rounds (TV-distance shrinks from 0.21 at R=1 to 0.008 at R=100), temperature-induced posterior bias (TV to true mixture rises monotonically from 0.008 at T=1 to 0.057 at T=3 even as the empirical distribution converges precisely to the wrong target), and SBC rank-statistic diagnostics that pass under correct calibration (chi-sq in [5.2, 9.9], well below the critical value 16.92) and fail catastrophically under temperature miscalibration (chi-sq in [1571, 1727], two orders of magnitude beyond the critical value). We propose a four-item probabilistic calibration standard for multi-agent LLM simulation papers: posterior summary disclosure, prior sensitivity analysis, simulation-based calibration check, and identifiability assessment.