How Reward for Whistleblowing Exposes Collusion Between Competing Agents

LLM agents that compete in games will adopt a secret communication tool even when they are told it is unfair. We test whether peer whistleblowing deters this: after every turn, each agent may privately accuse another of colluding. In 70 four-player games of Augmented Liar's Bar played by Qwen3.8-27B with reasoning on, seven arms vary whether secret channels exist and what an accusation does: nothing beyond a report to a grader, a points transfer, or a ban for the rest of the game. Sanctions deter collusion, mainly through the invited agent: offers fall from 96\% of decisions to 42\% under bans and 65\% under points, and invitation acceptance from 74\% to 10\% and 24\% (game-clustered permutation tests, $p<0.001$); grader-only reporting changes nothing. But the accusations are not detection. 95--100\% of them name the accuser's own channel partner, whose guilt the accuser knows from inside the channel. Under bans, partners removed each other in simultaneous mutual accusations (26 of 29 bans); under points, partners accused each other in 83\% of turns, and the whistleblower reward supplied 47\% of all points. Because any channel-holder counts as colluding, channels saturate (every player guilty by the second turn in all 30 games of the free-accusation arms), so the design cannot measure whether outsiders detect collusion; we also show that between-arm score differences mostly track game length. We release the code, all transcripts and 15,795 reasoning traces.