Plus Factors for AI Agents: Importing the Antitrust Standard of Evidence into Multi-Agent Collusion Audits

This paper borrows a good idea from antitrust law. Courts have known for a long time that firms moving in parallel is not proof of an agreement, so they look for "plus factors" - extra evidence that points to a deal. The authors turn six of those factors into measures that can be computed from agent logs, and add one of their own that only an AI auditor could use: replaying the episode with the communication channel cut, to see whether the outcome actually depended on it.

On their scripted testbed, where the true answer is known, the approach works well. Every single signal fails on some case, as the authors show clearly - high prices flag almost all of the tacit cases, a message judge flags most of the cheap talk, and price correlation actually points the wrong way. But their combined rule, using harm, the channel replay and one behavioural factor, catches 85% of cartels with no false positives in 60 confounder episodes. The most valuable part of the paper, though, is what happens next. When the same frozen rules are tried on real LLM agents that agreed prices explicitly in 10 out of 10 episodes, they catch none of them. The paper explains the three reasons honestly: the agents agreed prices below the theoretical Nash benchmark the harm measure relied on, the channel was cut at round 15, after the agreement had already settled, and a keyword judge did not recognise natural phrasing.

That failure makes the paper more useful, not less, and the revised suggestions are sensible: measure harm against what the same agents do without a channel, intervene early, and report harm and agreement as two separate verdicts.

Two suggestions: test the revised rule itself on fresh agents rather than leaving it as a proposal, and use an early channel cut, since agreements here formed within the first few rounds. A second judge, human or a different model, for the agreement labels would also strengthen the LLM part.

This is a strong paper that brings a new angle to AI collusion audits by borrowing the plus factors standard of evidence from antitrust law. It adds a seventh test that only AI auditors can run, i.e., replaying an episode with the communication channel switched off. The key finding is clear and convincing. LLM agents given a private channel explicitly agreed on prices in all 10 episodes, yet every one of the authors’ preset detection rules missed them, because the agreed prices were below the theoretical Nash price and cutting the channel at round 15 changed nothing. One obvious next step is to run more trials and test more models, especially frontier models, since the real agent results come from a single model (Qwen-Plus) with 10 runs per condition and no statistical test. The final audit standard, which calls for an empirical no-channel baseline and an early channel cut, could be another area for further exploration, since it was shaped after seeing the results and hasn’t yet been tested on fresh runs.