From the simulator, export respondent level choice predictions and assume the alternative with the highest share is the one chosen.
Now, let's say you had 3 holdout tasks. For each one that the prediction matches a respondent's choice, that's a hit. So if for a given respondent Jones your model correctly predicts one of the three choices, your hit rate for that respondent is 33.33%. If it was 3 hits or 0 then it is 100% or 0% and so on. Now average that across respondents. Voila, a hit rate.