Calling the Bluff: Forced-Deviation Audits of LLM Pricing Agents That Punish Slowly

Pricing agents built on large language models generally converge to keeping prices high without being told to collude in supracompetitive environments, and auditors need a way to tell when collusion occurs. Tran and Yu (2026) proposed a forced-deviation audit in which they ran a market scenario where they forced one agent to cut its price and checked whether the rival punishes it in the next period. We show that this one-period test can miss agents that may change slowly, or respond with stickiness. A scripted rival that punishes but moves only 8% toward its target price scores +0.086 on the one-period test, indicating no punishment. We propose a dynamic baseline that predicts how a slow rival that does not punish would respond, and we score the rival against it. The same punisher scores −0.571, and a slow non-punisher scores zero even with the cut persisting over periods. When Claude Haiku 5.5 is instructed to punish cuts, the one-period test misses it in all four usable forks, because the agent acts one period late, while the dynamic measure detects it in all four. Auditing uninstructed Haiku 5.5 (12 runs, 48 cuts, ρ̂= 0.92), the one-period test finds no punishment (+0.057, 95% CI 0.049 to 0.065), while the dynamic measure finds a response below the non-punishing baseline (−0.534, 95% CI −0.856 to −0.194) that is specific to cuts but too weak to make cutting unprofitable. This experiment further highlights the importance for audits of capturing how slowly an Agent moves before punishment.