Unsupervised risk estimation — partially specifies error not parameters → Partially specified models

Unsupervised risk estimation is explicitly named in the text as an instance of partial specification, but it partially specifies a different target than the method-of-moments examples that motivate the family. Where GMM-style partial specification assumes something about the noise or moments of a regression in order to identify the model's parameters, unsupervised risk estimation ignores the parameters and the rest of the data distribution entirely, positing structure only in how the model's errors are distributed, sufficient to estimate risk on unlabeled test data (concrete-problems, §"Partially specified models: method of moments, unsupervised risk estimation, causal identification, and limited-information maximum likelihood.", p. 18). The paper notes this narrower target is also what gives it reach: because the goal is only to output a large risk estimate when things have gone badly wrong, it can in principle handle test distributions too different from training for accurate prediction to even be a coherent goal.