How do we tell coordination apart from shared causes?

The abstract says the paper will "specify the evidence needed to distinguish influence from common causes." The excerpt gives the commitment and not the list.

The reason it is needed is that similar behavior has two sources. One agent may have influenced another, through a transfer. Or both may have been pushed by something they share, such as the same model, the same instructions or the same environment. A monitor that sees two agents doing the same thing cannot tell which without some causal model. Two illustrations of mine, not the paper's: two agents return the same answer to a web-retrieval task because the same page exists, and two workloads find the same weakness because it sits in the environment both were placed in. Either could look like sharing.

The vault has both halves of this problem in other places. Does receiving misaligned email cause agents to send it? removes one common cause, agent-level differences, and Does receiving misaligned email cause agents to send it back? shows what is left open. Does peer behavior actually cause collusion between agents? is the vault's one reported intervention on a peer, in a two-agent collusion task, and its excerpt does not say what was manipulated. The lever also differs from the ones proposed here, a peer's behavior and not a channel or a store, so it shows what an interventional answer looks like and does not answer the defence-side question. Does a multi-agent setting automatically signal a security effect? asks for a baseline before crediting interaction. Here the same discipline is applied on the defence side, where a false attribution costs reviewer time and a missed one costs an intrusion.

What could count as evidence is open. A record of a transfer, where one execution wrote something a later one read, is observational. Closing a channel and seeing whether the behavior returns is interventional, and the proposed evaluation "tests recurrence after channel closure and state quarantine" (Does added monitoring improve protection at acceptable cost?). That the recurrence tests serve this separation is my reading, and the excerpt does not link them. The vault states the confound in the opposite direction for validators: Can a quorum of validators really provide independent judgment? says a quorum's agreement can come from a shared cause and not from independent confirmation, and Does model diversity actually reduce validator agreement failures? proposes varying one shared channel at a time, an intervention of the kind this paragraph lists as open.

What the excerpt does not give. The evidence the paper specifies, whether it is observational or interventional, and any case where a common cause was ruled in or out.