In this context, we probably don’t get isolated outcome events with precise measurements as our ground truth. A given test has a precise score, but the key measures of success are longer term performance with credit assignments & assessments that are more vague & subject to debate. To make progress it is probably necessary to make those best-effort assessments – what we are trying to optimize along with multi-attribute costs – as ground truth/parameters of a given model. The optimization goal is to figure out how to optimize interaction with & investment in pools of candidates in order to get the optimal estimated outcomes relative to a fixed system of performance assessment taken as a situational parameter.