Same Prices, Different Policies: A Control Suite for Cost-conditioned Algorithmic Pricing Audits

Price co-movement, high margins and retaliation are different observations, but a collusion audit can blur them together. We build a reproducible synthetic control suite that keeps those observations separate. Five constructed policy pairs and 32 independently trained, cost-conditioned Q-learning pairs face matched common-cost paths in a finite logit pricing game. The controls include two pure Nash equilibria, myopic best responses, nonreactive high-margin pricing, and a hand-coded grim-trigger positive control. A price-correlation screen alarms on every one of 100 held-out paths for each competitive/myopic control. A conjunction of a cost-aware profit index and a paired unilateral-deviation response instead alarms on none of those controls and all grim-trigger controls. The nonreactive high-margin control demonstrates why that conjunction is not a general test for collusive outcomes. For learned policies, the mean within-policy combined alarm rate is 47.1% (95% policy-cluster bootstrap interval 42.5-51.4%) in the primary setting. It falls from 65.9% to 25.5% as the response horizon increases from four to 24 periods. We also enumerate grid equilibria and show that choosing the lower equilibrium as the competitive reference can assign the higher equilibrium a profit index above the alarm threshold. These are calibration results, not validated detection accuracy, evidence of intent, or results about LLMs or real markets.