Say, Do, or Both? Catching Collusion Through Cross-Agent Consistency

Context and motivation. LLM agents increasingly bid and negotiate for firms. When two collude, a monitor can read what they say to each other or what they do in the market, and the two channels disagree. Question and problem. Does reading both channels beat reading one, and what counts as evidence of collusion rather than talk, coincidence or a one-sided favour? Principal ideas and results. We built a repeated procurement auction with three LLM bidders, two sharing a parent company, and 48 labelled games across six case types, each designed to fool one channel. A frozen LLM judge read messages, bids, both, or a structured report; we added bid screens, a cross-agent consistency check and a small classifier. A judge reading both transcripts did not beat the best single channel (∆AUROC +0.02, p = 0.33), inheriting the bid channel’s false alarm on innocent rotation. Cross-agent consistency flagged no innocent game. The classifier did beat it (AUROC 0.99, ∆ +0.10, p = 0.003), its gain coming from bid features. Unscripted, a group bonus produced a cartel in round one unless the prompt stated a law or a watching regulator; across six further model families this gate is model-specific, and some collude under the law and declare it. The monitors’ ranking transfers (classifier AUROC 0.87), but frozen thresholds catch about half the new cartels. Contribution. A reusable benchmark with known blind spots, a formal test of channel combination, and detection signals whose ranking holds across model families.