That implies an exclusively pessimistic model, however. My suggestion was to simply have a more pessimistic prior. As you say, insurance has to pay out _something_ _sometime_ by the data, and I've certainly seen some of those payouts myself (and sometimes even had them be advantageous in my favor), but if there's enough prior for adversarial interaction it seems a more more robust approach to plan for a worst case scenario even if it doesn't end up happening; or at least don't be exclusively optimistic, for this provides the same but inverted issues as exclusive pessimism.