The Covert Communication Frontier: Channel Capacity, Harmful Pricing and Auditability in LLM Markets

LLM pricing agents could coordinate through side channels that carry no payoff. We ask how much harm such a channel enables as its capacity grows, and whether an auditor who sees only market data can detect it. Two Qwen2.5-7B sellers play a repeated Bertrand game, seeing each other's prices with a three-round delay, plus a binary timing channel whose capacity $q$ is the probability that a bit is delivered rather than replaced by a decoy. Given a fixed convention (scripted sender, LLM receiver told the convention), harm rises monotonically with capacity, the final window collusion index goes from 0.58 at $q=0$ to 0.88 at $q=1$, and harmful episodes from 1/10 to 10/10 (Wilcoxon $p=0.004$ on per-seed slopes). Agents did not discover the channel unaided, unprompted senders sent a constant bit, scaffolded pairs fell into price wars in 2 of 3 runs, and prompted LLM senders could not hold a 40-round signalling rule, although the LLM receiver followed signals reliably. A market-only auditor detected harmful episodes above chance (AUROC 0.70, 95\% CI 0.55--0.84); timing bits added nothing, because here coordination shows up in prices. Channel capacity is a measurable lever for limiting harm, and at 7B scale committing to a convention is the binding constraint.