TwinSignal: observational twins and randomized channel probes for agent audits

Correlated actions can reflect either communication between agents or a shared environment. Treating statistical dependence as proof of collusion therefore risks accusing agents that never exchange information. We construct an executable synthetic benchmark in which a common-cause mechanism and a directed-copy mechanism produce exactly the same passive action traces. No statistic computed solely from those traces can distinguish the mechanisms. We then randomize the signal delivered through the candidate channel and measure the recipient's response using a one-sided exact Fisher test. A factorial experiment varies sample size, response noise, and intervention uptake across 400 seeds per configuration. With 128 rounds and 10% response noise, passive association is detected in all 400 common-cause and all 400 directed-copy runs. Under full channel intervention, the corresponding rates are 9/400 and 400/400. When intervention uptake is zero, the two mechanisms again coincide. These are controlled binary-policy simulations, not observations of language models or unlawful market conduct. The contribution is a small, reproducible audit fixture with matched negative and positive controls and an explicit failure condition. It demonstrates why an audit should separately report action association, channel responsiveness, and evidence of harmful coordination, rather than treating them as interchangeable findings.