One thing worth having before the Steam demo, if it is not there already: a headless sim that plays a naive greedy line over a few hundred seeds and reports the distribution of outcomes rather than a pass or a fail.
Two traps we hit doing exactly this on our own balance work. First, count the outcome you actually mean instead of deriving it. We had a loss defined as not a win, and that hid the fact that most of them were stalls rather than defeats, which is a completely different fix. Second, put a ceiling on the win rate and not only a floor. A floor catches the build being too weak, and drift almost always goes the other way once players find the strong dice combination.
For a dice game the useful readout is usually which turn the run stopped being in doubt, rather than whether it finished.