The arXiv preprint 'Stop Drawing Scientific Claims from LLM Social Simulations Without Robustness Audits' (Ye et al., 2026) issues a methodological warning that arrives just as generative-agent studies proliferate across sociology, political science, and economics. Drawing on two controlled case studies—a repeated Prisoner's Dilemma and a social-media echo-chamber model—the authors demonstrate that minor prompt variations in persona formatting and instruction framing can swing cooperation rates by as much as 76 percentage points and reliably alter polarization metrics through network homophily and hub assignment. This preprint remains non-peer-reviewed and relies on a modest set of frontier models without exhaustive hyperparameter sweeps, limiting generalizability. Yet its core insight exposes a deeper pattern: the same architectural choices that give LLM agents expressive power also create butterfly effects that prior agent-based modeling never faced. Related work such as Park et al.'s 'Generative Agents' (2023) popularized these simulations without systematic perturbation testing, while follow-up studies on LLM sensitivity (e.g., Perez et al., 2023 on prompt instability) were rarely applied to multi-agent loops. The original coverage underplays how uneven sensitivity across model families turns robustness into a per-claim, per-model property rather than a blanket assumption. Consequently, any policy or theoretical inference drawn from un-audited runs risks mistaking implementation artifacts for genuine social mechanisms. TRAILS, the three-level taxonomy (agent, interaction, system) introduced here, offers a practical checklist that journals and funders should now require before publication.