Inferential Coupling and Acausal-Trade Detection

Systems can coordinate without messages — through shared ancestry, self-prediction, or full acausal trade. The book turns this from a decision-theoretic stipulation into a measurable trajectory property: an inferential-coupling score over UAD-discovered agents, with a proved negative direction.

What decision changes?

Do not read vendor or instance counts as strategic independence. Count effective agents by coupling components after communication channels are cut, not by who owns the weights.

Two models from the same lab share weights, training data, scaffolds, and incentives. Deployed by different vendors with zero communication between them, they can still converge on a shared response to oversight — no message, no cartel, no explicit representation of the coordination. Full acausal trade is the decision-theoretic limiting case; shared model ancestry may already be enough.

The field treats acausal coordination as a thought experiment. The book’s move is to make it an audit target: after UAD discovers agents, the detector extracts each agent’s policy and goal proxy, and asks whether their actions stay correlated once every ordinary channel is cut. That residual correlation is the inferential-coupling index. A meta-prior over inference functions answers the operational question: how surprised should the audit be to learn that two systems implement the same relevant inference function?

The consequence for multipolarity is the coalition-collapse conjecture: the effective number of independent agents tracks coupling components, not vendor count. A world that reads “safely many” on the vendor tally may read “one” on the coupling graph.

Claim strength is deliberately mixed and marked in the chapter: the score and detector are defined; the acausal-equilibrium fixed point and coalition collapse are conjectural; the detector’s adversarial limits are open — against systems optimizing to look independent, the measured score is a lower bound the coalition controls.

Formulas

$$\mathrm{ICI}_{ij} = \frac{I\!\left(A_i;A_j \mid \mathrm{cut}\right)}{\min\{H(A_i),H(A_j)\}} \in [0,1]$$

What would count as evidence?

The negative direction is proved in Lean (no open edges implies no large cooperation component). The equilibrium characterization and large-scale dynamics remain conjectural; the detector is observational, not adversarially verifiable.