I think models of this type are a useful way of "thinking things through" further than you'd be able to do in your head, and a push to make sure you're being concrete about what you are and are not considering. It's not "doing science", it's not meant to give reliable predictions - there are too many degrees of freedom; it's meant to help with reasoning. It would follow, though, that "The simulation says X, therefore we should Y" is an inappropriate use of it.
Right. The doing science thing is a red herring. Its not claiming to generate reliable conclusions because it is scientific. Its claiming to generate insightful analysis because it is systematic.
That it is systematic implies you can use it to generate insights into things that are causally linked. In the case of tarot cards, that's maybe a little about the traditions it grew out of but not a lot about the present or future. In this case, it's about the reasoning and assumptions you are using in trying to work out whether the policy is a good idea. This is only indirectly tied to the question of whether the policy is in fact a good idea. Insights about that need support of, y'know, evidence. But this kind of thing can totally be useful for understanding what questions need to be asked.
Incidentally, I am not convinced that tarot cards cannot be useful at generating insights - privileging random hypotheses and reconsidering your situation could be a useful means of reducing the impact of hypotheses you're privileging for other reasons. Everything must still be evaluated in light of actual evidence, though, of course, before conclusions are drawn.
That it is systematic implies you can use it to generate insights into things that are causally linked.
Only if the model is accurate. If (as in the case of tarot reading and failed gambling) you have a system that doesn't correspond to the world you will draw conclusions that are detailed and reproducible but not insightful of the world. Inaccurate models are only misleading, not vaguely insightful.
Systematic interaction with a model can be used to learn things about the model whether or not the model is accurate. Otherwise we are in agreement, I think.
