My thoughts are always around the concept of having layered cause-and-effect. That is, the scenario (universe) has some degree of linearity in terms of how (for example) the 'war' gets started, and how it ends. Ummm, which is a little vague.
One idea, specifically, I had was an FPS/Flashpoint style game set in an occupied and ruined city; with a world extending to a few kilometers of countryside outside (at which point they'd hit an insurmountable level of opposition, such as a hunter-killer chopper, because you can't have an infinite world to play in). The players' objective would be simply to 'liberate' the city; the enemies to occupy it.
Each side - rebels and occupiers - would have a specific set of needs; food, arms, territory. These needs would be fixed rules. Below the 'needs' would be 'tactical needs'; these would be more randomised, but also fixed to certain strictures - for example, to hold position XXX for a food convoy, raid position YYY to hunt rebels, destroy occupier position YYYY. Tactical needs would be grouped based on which need they fulfilled; the AI (determining the dynamic battlefield) would determine the importance of needs based on which were strongest and which were weakest; for example the rebel 'command' AI would focus on food raids of enemy convoys or bases if food was in short supply, or liberating territory if it was losing ground.
(these needs would also be finite resources, such that shortage would force retreat and potential defeat)
Tactical needs would, naturally, exist in multiples; they'd be performed both by AI simulation and player action, although the 'feed in' to the simulation part would be influenceable by player action. For example, if the player has succesfully destroyed several tonnes of food in an enemy position in prior missions, then the opposition troops will be weaker there (in the simulation calculations). It'd also be consequent that the players' success is exaggerated (the player does well, the 'war' goes well) but their failures are not (to avoid a spiral of doom) - but only in a manner perceived to be realistic.
Anyways, the tactical needs would by dynamically broken down into the actual missions; missions that had several objectives as mentioned prior. But this environment would be stochastic; defenses would be changed on the basis of AI, randomisation and prior events (such as weapon shipments or even prior attacks). Objectives would have their positions varying to literally anywhere logical. Say you have a raid to capture a map; that map could be on the body of a dead soldier, on a table, in a truck, or thrown on the fire when you attack; each of these events is believable yet random. At the same time, each objective is non-boolean; you can do xx% of it, so success is measured in terms of grey - the player has to judge when risk is acceptable (accounting for losses) based on the reward.
Most games do this anyways, but not in terms of risk to the 'world', only risk to the player; 'tactical retreat' should be a viable tactic, and better than a phyrric victory in this sort of simulation.