Distinguishing Governance Effects from Prompting Artifacts in LLM Pricing Simulations

We study a small-scale LLM-based social simulation of repeated algorithmic pricing in a synthetic three-retailer market. Repeated interaction means that each agent reshapes the simulated market environment across rounds, allowing us to assess whether natural-language institutional constraints are associated with subsequent behavior. The simulation compares a profit-oriented baseline, a generic managerial-discipline placebo, and a governance-policy condition using DeepSeek-V4-Flash across 12 seeds per condition and 8 rounds. The observed governance-policy pattern differs from both controls in a placebo-controlled comparison intended to distinguish rule-specific content from generic prompt artifacts. Relative to baseline, the governance-policy condition was associated with lower violation rate (0.281$\pm$0.121 to 0.219$\pm$0.152), coordination-follow rate (0.250$\pm$0.131 to 0.156$\pm$0.108), and focal-retailer price volatility (0.0568$\pm$0.0363 to 0.0263$\pm$0.0115), while mean market price was nearly unchanged (2.887$\pm$1.028 to 2.881$\pm$1.029) and focal-retailer profit retention was 0.9815. The placebo condition provides a comparison for distinguishing substantive institutional content from generic managerial wording.