PRISM: What Would People Do? Evaluating Demographically-Stratified LLM Agent Populations as Models of Human Response to Interventions

Predicting how a diverse human population responds to an intervention — a policy change, a framing manipulation, a market shock — is a general problem across the behavioral sciences, yet it is poorly served by statistical models that capture correlations without mechanisms. We introduce PRISM (Population Response and Intervention Sensitivity Model): a framework in which a population of LLM-backed agents, parameterized by empirical demographic distributions, generates grounded sequential decisions and is evaluated for behavioral validity along four dimensions — population-level realism, mechanistic explainability, demographic heterogeneity, and intervention sensitivity. Validated externally against published human data, PRISM reproduces the reported distributions and treatment effects of eight canonical experiments (framing, ultimatum, dictator, anchoring, conjunction, trust, default, and endowment effects) with a mean absolute error of 4.0 percentage points ($r = 0.98$ across 15 measures), including three intervention-driven reversals (framing, default, endowment) that a stereotype-only model cannot produce. Instantiated in depth on one economically important special case — subscriber response to market shocks in a streaming service — it reproduces segment-specific elasticity patterns, extracts causally verified decision pathways, and reduces calibration error against published elasticities from 32.5 to 3.0 percentage points relative to a survey-style LLM baseline. Our contribution is a validated methodology, not merely another simulation: the four-dimensional criterion tells a practitioner whether to trust a given agent population before acting on it. PRISM thereby reframes population behavior modeling from correlational prediction, which cannot extrapolate to never-observed interventions, to mechanism-generating simulation, and proposes intervention-driven reversals — cases where the population is held fixed and only the environment changes — as a discriminating test that separates genuine behavioral mechanism from demographic stereotype.