Maybe training with actual bottom-line utility function would have shown some use case, that is, instead of MCC you would have predicted whether sending a perso

Maybe training with actual bottom-line utility function would have shown some use case, that is, instead of MCC you would have predicted whether sending a person to change the drive now before it fails would cost less than letting it fail and cleaning up afterwards. But if the drives are part of redundant array, it would be almost always cheapest to let them fail.. and if large part of failures are asymptomatic, you need the array anyway for critical stuff, so I suppose it's a useless exercise.