Studies of human conversation suggest that agents whose world models are in consensus can work well together using only very narrow bandwidths. The total bandwidth required between agents could hence be minimised if we could recognise when model consensus breaks down. At the breakdown point, the communication policy could switch from some usual-case low value to a temporary high value while the model conflict is resolved. To effectively recognise the breakdown point, we need tools that recognise model conflicts without requiring extensive bandwidth. A mathematical model of probing and-or graphs suggests that, for a large range of interesting models, the number of probes required to detect consensus breakdown is quite low.