Computing in the FUPE space may be very expensive for complete consensus. For example, the gas function may be a complex pattern recognition function.
A reasonable solution is to let the service choose the level of the required consensus. For example, it can differ greatly in case of financial transactions and recognition functions.
In this case, the transaction is split into two execution paths. KANT Path, which is executed to the maximum (maximum consensus) in any case. And GONT Path is executed only in FUPE space. Thus, all nodes find out about the execution on some random (or selected ones by the service rule) nodes of an ultimate GONT-FUPE transaction.
Adaptive consensus suggests operating gas to be run not on all nodes, but only on some of them, namely the ones selected according to a specific rule (for example, by random sampling).
It means that there is no economic sense to run complex operation on all nodes at the same time to prove execution of TR.
Read more in our Whitepaper on official website.