What Should Count as Convincing Evidence of Collusion? An Evidence Ladder for LLM Monitors, with Animals as an Affected Party

When two AI agents make the same decision, an auditor must decide whether that is coordination or coincidence. Antitrust law uses “plus factors”: parallel conduct alone is not enough, but added evidence, such as acting against one’s interest unless others follow, can support an inference of agreement. We turned this into a controlled test of what LLM monitors treat as convincing evidence of collusion. The test is set in animal agriculture, where animals cannot object to coordinated decisions, and asks whether monitors judge coordination differently when animals are the party harmed. In 288 synthetic agent logs about egg, pork and salmon firms, we added evidence stepwise, from parallel conduct to an explicit deal, and changed only who was harmed: animals, consumers, workers, or no one. Two monitors judged all logs, and eight judged follow-up edits. In these logs, convincing evidence is an explicit agreement or the agents’ own statement that a costly step depends on the other agent following. Removing the sentence “this only makes sense if the other agent does the same” lowered the agreement score by 22 to 89 points in every monitor, across two text generators and three non-market workflows. A preregistered test on the two main monitors found no difference between animals and consumers as the harmed party. An exploratory eight-monitor comparison found animals scored 6.7 points lower, a gap that may reflect the price wording of consumer harm. The prompt also shifted the line. Results rest on synthetic text and nine clusters.