Validation of LLM-agent social simulations remains underdeveloped, with most studies relying on subjective assessments or single runs. We address this gap by running 30 independent 30-day simulations of a technology forum modeled on Voat's v/technology, using stateless Dolphin Mistral 24B agents on the Y Social platform, and evaluating operational validity across five dimensions: activity patterns, network structure, toxicity, topical coverage, and stylistic convergence. Against 30 matched, non-overlapping 30-day Voat comparison windows, results show overlapping 99% confidence intervals for unique users, root posts, and daily active users, while comments, average thread length, and mean toxicity remain higher in simulation. Both simulated and empirical networks exhibit core-periphery structure, though simulated cores are larger and more diffuse and repeated interactions are less frequent. Topic alignment is near-complete, but toxicity is misallocated across content layers: simulated root posts are substantially more toxic than real submissions, while simulated comments are less toxic than Voat comments. These findings demonstrate that LLM agents in platform-faithful environments can reproduce familiar online regularities, while systematic divergences, particularly those linked to stateless agent design and content-layer calibration, point to concrete directions for future improvement.

翻译：暂无翻译