Plus Factors for Collusion Monitors: LLM Monitors Mistake Open Cooperation for Collusion

LLM monitors are increasingly proposed to catch AI agents that collude, but agents that cooperate openly can look a lot like agents that collude. I introduce Séance, a four-player LLM guessing game in which a secret pair may pass information through public messages while honest players often share their information openly. On a held-out test set played by Qwen3.8-27B, I find that every LLM monitor I tested, from Qwen3.8-27B to Claude Opus 5.5, accuses most honest games when given a plain prompt, even though that prompt describes collusion as covert. Opus accused all 50 honest games while naming the colluding pair in every collusion game, which suggests that spotting coordination is easy and judging whether it is collusion is hard. Borrowing the "plus factors" that antitrust courts use to tell agreement from parallel conduct, I add one paragraph to the monitor's prompt that says open sharing is not evidence and lists what is. Without labeled data, it cuts false accusations for nearly every monitor (Qwen3.8-27B, 43 to 19 of 50), though most monitors also miss more real collusion. Switching model families or using calibrated decision models does not fix the problem. This failure stays invisible unless a monitor is tested on honest cooperation, so collusion monitors should be checked against such cases before anyone trusts them. I release Séance as the first game of a proposed benchmark of these hard negatives for collusion monitors.