A method and associated systems for forecastable supervised labels and corpus sets for training a natural-language processing system. An NLP-training system asks an “oracle” expert to answer a predictive test question and, in response, receives from the oracle an answer, rationales for selecting that answer, and identifications of extrinsic natural-language sources of evidence that supports those rationales. The system retrieves updated versions of that evidence at a later time, and returns that updated evidence to the oracle. In response, the oracle returns an updated answer and rationales based on the updated evidence. The system then compares time-varying characteristics of the evidence in order to determine the relative contributions of each piece of evidence to the oracles' selections. Less relevant evidence is discarded and the remaining, optimized, evidence is forwarded to the NLP system to be used as training data.