Speaker
Description
Learned network digital twins promise what-if analysis and safe policy training without touching production networks, yet they are typically validated by how accurately they reproduce observed behaviour. Accurate reproduction does not establish that a twin answers counterfactual questions correctly, that it encodes the physical properties of the path, or that controllers tuned inside it still perform well on the network it models. We propose an evaluation protocol that measures these properties separately. In a single-bottleneck congestion control setting, the sender observes only acknowledgement-derived telemetry, while link capacity, buffer depth, base round-trip time and cross-traffic are hidden but fully controlled, and any what-if scenario can be replayed exactly from the same seed. We assess twins built from recent world-model architectures on four axes: prediction accuracy over horizons measured in round-trip times; recovery of hidden path parameters, compared against minimum round-trip time and packet-pair estimators; agreement with the simulator on the effect of alternative sending rates; and the performance of rate controllers trained entirely inside the twin. We hypothesize that prediction accuracy is a weak indicator of the other three, and that controllers trained inside a twin exploit regions where it is optimistic, such as underestimated queueing delay. Initial experiments use a fast queueing simulator, and the simulator and evaluation suite are released for reuse.