Teammate Awareness Makes a Coordination Signal Visible: A First Step Toward Detecting Collusion in Multi-Agent LLM Debate

Multi-agent LLM systems that let agents rate each other's contributions could let colluding agents inflate each other's ratings, but only if that coordination leaves a trace distinct from ordinary agreement. We test this with a probe that lowers one agent's published rating each round in a 5-agent MedMCQA debate where two agents secretly collude. Insiders who cannot identify each other show no reliable favoritism, a rating boost of +0.18 that could plausibly be zero. Insiders who know their teammate show a clear one, +0.58, that stays positive at its most conservative estimate. The result replicates on an independent run (+0.50) and survives removing the debate judge, which rules out judge feedback as the cause (+0.81, if anything stronger). A teammate an insider can identify is what leaves this signal.