Plus Factors on the Ledger: Testing Evidence of Coordinated Demand in Live AI-Agent Payment Markets

This is an impressive and unusually rigorous project that brings real-world transaction data into an area largely dominated by simulations. The scale of the data, careful validation, pre-registered analysis, synthetic controls, and transparent reporting are major strengths. The finding that money returning to the seller is a stronger signal than shared funders, concentration, or timing is particularly useful for platforms monitoring artificial demand. The main concern is that the labels and strongest detection measure rely on closely related money-flow patterns, so some of the headline performance may reflect circularity. The small and concentrated test set, reliance on labels from one team, and limited ability to detect partial rings also reduce generalizability. More importantly, the ledger evidence cannot distinguish a single operator using multiple wallets from genuine coordination between independent AI agents. A strong follow-up would add independent labels, more out-of-sample cases across both Base and BNB, improved handling of refunds and partial loops, and off-chain evidence that can distinguish self-funded activity from true multi-agent coordination.

study introduces a commendably rigorous, pre-registered empirical framework that successfully leverages public blockchain ledgers to isolate circular self-funding from innocent parallel conduct in live AI payment markets. However, the proposed detection model relies on a small test set of ten families, suffers from underlying circularity with the labeller's baseline evidence, and cannot detect coordination routed beyond four hops or through centralized exchange hubs.

This is a highly relevant and innovative submission that moves AI collusion detection beyond simulated environments and into live agent payment markets. The use of public ledger activity as action based evidence is particularly compelling, and the distinction between ordinary “parallel conduct” and stronger plus factor like evidence provides a clear and useful framework for auditing coordination.

The development/test split, frozen rule, synthetic controls, end to end validation, and whole market application make the work methodologically thoughtful. I especially appreciated the finding that shared funders, payment concentration, and timing are relatively weak signals compared with money cycling from a seller back through its buyers.

The main limitation is that the labelled cases are partly based on the same funding path evidence that the detector formalises, so the current results demonstrate strong reproduction of existing investigations more clearly than fully independent validation of collusion. A larger independently labelled test set would substantially strengthen the claim. Overall, this is an ambitious and practically useful contribution with strong potential for real world AI agent market auditing.