[Micah]

This is probably going to sound weird, but how about if the game processed the combats multiple times? The first time through the combat would be run using gems and executing retreat orders. If the battle is easily won (less than 10% casualties and no commander deaths) the battle would be rerun without the use of gems or the execution of retreat orders. If the gem-comservation run results in a large swing in the outcome (more than the 10% casualty threshold or commander death) the results would be discarded, otherwise they would be used.

By only rerunning battles that are lopsided to begin with this method wouldn't be giving any "second chances" to land/avoid a lucky blow. The results of a battle will never be changed using this method, only the resources that go into winning it.

This gets rid of the need for the AI to have to perfectly assess the strength of the opposition with a complex formula that will inevitably have holes in it and lets it see what actually happens. It's a bit unthematic that mages can never be surpried by a small force with big guns, but eliminating the frustration of dealing with the AI's flaws seems like a very worthwhile tradeoff to me. I'm sick of seeing my mages flee in terror from a scout that my PD caught when I have an uncastled lab, or the mage I have around to buff an SC and split flee from a pack of overgrown birds, leaving his SC buddy around to die when the real army gets there later in the turn.

Single-target charm-type spells should always be cast as scripted though, since getting a unit is even better than not losing one, and since they single-target there is no chance of overkill.

It might up the turn processing time a bit, but since only lopsided battles will be rerun (which generally process fairly quickly, with the exception of SCs taking down armies) I don't think it would be a major resource-hog.