Are you sure that you don't first look at the behavior and then calculate an incentive map?

Well, both. Normally I estimate (and update) the model(s) in the middle of an interaction. Before I have no data and have to fall back on priors, and after I have no need for a model.

Are you saying there are, um, methodological problems with this approach?

doesn't your last paragraph depict a thought process which is the exact opposite of Hanlons razor?

Doesn't look like that to me. The opposite of Hanlon's Razor is "I don't understand her therefore she is trying to hurt me". I'm starting by trying to figure out what the person wants and only if I fail I start to consider that she might be clueless (as Hanlon's Razor would suggest) or mean (in case Hanlon's Razor is wrong here).