FedCausalCompose Shows Observational Models Incur Irreducible Interventional Error Under Unblocked Back-Door Paths

The paper introduces FedCausalCompose, a framework that treats cross-module API calls as interventions to recover causal interfaces between order, payment, inventory and shipment services. Local action-response pairs supply the data needed to distinguish authorization edges from mere temporal precedence. This directly addresses the gap between observational traces and intervention-time validity in composed agent systems.

Diagnostic experiments isolate two conditions. In structured tool environments with explicit preconditions, causal composition outperforms non-causal baselines once interface coverage reaches levels that close back-door paths. In dialogue and narrative settings, raw edge lists produce no measurable gain unless an attention anchor renders the causal information usable at decision time. These results quantify the statistical identifiability requirement stated in the abstract.

Prior agent work on world models, such as the Voyager paper and ReAct traces, relied exclusively on observational fit and therefore inherits the same interventional bias documented here. The new framework supplies the missing do-calculus step for modular systems. It also reveals why current benchmarks that score only trajectory match fail to predict deployment reliability.

Next steps require controlled release of intervention logs from production tool suites to test whether the coverage threshold generalizes beyond the diagnostic environments. Agent training pipelines that embed causal interface discovery as a separate objective can then be evaluated against the non-causal lower bound established in the paper.