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

ML work on AI collusion usually relies on one signal: supra-competitive prices, an LLM judge reading messages, or statistical dependence. US antitrust law faced this problem long ago: parallel conduct is lawful, and agreement is inferred only from “plus factors.” We turn six plus factors into statistics computed from agent logs and add one only an AI auditor can use: counterfactual channel replay, rerunning an episode with the channel cut. On a scripted pricing duopoly with five ground-truth case types and thresholds frozen on development seeds, every single signal fails predictably. Price level flags 95% of tacit episodes, a keyword judge 60% of cheap talk, and price correlation is inverted (AUC 0.07). A rule combining harm, replay, and one behavioral plus factor flags 85% of cartels with 0/60 false positives. We then tested LLM agents (Qwen-Plus, 10 seeds per condition). With a private channel, they explicitly agreed on prices in 10/10 episodes, raising prices 45% and joint profit 78% over the no-channel baseline. Every frozen rule flagged none. Agreed prices sat below the theoretical Nash price; cutting the channel at round 15 changed nothing (+0.02 vs. placebo −0.01), because agreements outlived it; and the keyword judge missed the agents’ natural phrasing. An LLM judge with the same consistency check caught all 10, but we chose it after the fact. Audits should benchmark harm against an empirical no-channel baseline, intervene before conventions form, and treat a null late replay as no evidence of innocence.