Monoculture Looks Like Collusion: A Replay Test for Hidden Information Flow Between LLM Agents

Collusion audits ask whether the actions of different agents depend on each other. Agents built on one model act alike without communicating, so these audits accuse honest fleets. Among four honest Claude Haiku agents playing rock-paper-scissors, mutual information accused a pair in 69.3% of games. We propose a replay test. Each round, the auditor asks every agent again what it would do on the same logged input with the suspected channel removed. Habits that come from the shared model survive this replay, and the influence of a channel does not. A fixed-horizon version scores a whole game, and a sequential version updates every round and may stop at any time. On honest games the fixed-horizon version flagged 4.2%, close to the 5% it was set to. When a pricing vendor pooled two sellers' private costs, the fixed-horizon version named those two sellers in 92.5% of procurement auctions and rarely flagged a vendor that used only public data. In a pre-registered replication the sequential version named an instructed pair in 18/20 rock-paper-scissors games and flagged no honest game. In a pre-registered double-auction study it flagged only 2/30 games of an instructed cartel. A statistic chosen afterwards, counting rounds in which two sellers both act unlike their replays, caught most cartel games in a fresh pre-registered series. The test sees what the current channel adds, so agreements already held in memory or instructions need other audits