When Do AIs Have a Meeting of Minds? Plus Factor Evidence of LLM Agent Collusion

This is a really interesting project in that it directly tries to track the current existing legal framework that tries to infer when illegal collusion has happened in the absence of explicit agreement, that is "plus factors."

This is a conceptually original and well-executed submission that effectively connects antitrust doctrine with ML-based collusion detection. The three-tier evidence framework is practical for regulatory investigations, and reporting false positives alongside detection rates is a notable strength. The study also demonstrates strong research discipline, reproducibility, and transparency. However, the A vs. B comparison may be confounded by communication availability, making A vs. D a stronger primary comparison. Tier 3 also requires further testing for potential label leakage by masking message-related information from private reasoning. The relatively small control samples and reliance on a single human coder are additional limitations. Expanding the control data, adding an independent coder, and testing agents that conceal reasoning would strengthen the results. Overall, this is an innovative and relevant paper, and addressing these methodological concerns would significantly improve its credibility and enforcement value.

This paper takes the same antitrust "plus factors" idea and asks a sharper question: not whether a market looks coordinated, but what evidence would justify saying a particular seller actually agreed. It reanalyses 75 published market runs, sorting the evidence into three tiers — public conduct, the sellers' messages, and their private reasoning — and labels agreements using a codebook grounded in US case law, applied by a rule-based labeller that agrees well with hand labels.

The results give investigators a clear order of trust. Public conduct alone catches about three-quarters of agreements but also flags 29% of sellers who could not communicate at all — far too blunt to accuse an individual seller. Private reasoning is much sharper, separating agreement from parallel pricing almost perfectly. But the paper is careful about what that reasoning score measures: intent, not agreement. It flags 40% of sellers whose invitations nobody accepted, and it misses agreements formed silently, through conduct alone. The judge's plain yes-or-no verdict is also shown to be unusable without calibration — it says "coordination: yes" in 69% of hours where sellers could not even talk.

The conclusion is well earned: only a recorded invitation and an acceptance really show a meeting of minds, and conduct and reasoning should be treated as pointers for where to look, not proof. The practical recommendations follow naturally — keep message and reasoning logs, and calibrate any AI judge against a no-communication baseline before acting on it.

One suggestion: the false-positive rates rest on only five no-messaging markets per half, so testing the frozen rules on a larger set of innocent markets would firm up the numbers investigators would rely on most.