Procedural content generation via machine learning (PCGML) is typicallyframed as the task of fitting a generative model to full-scale examples of adesired content distribution. This approach presents a fundamental tension: themore design effort expended to produce detailed training examples for shaping agenerator, the lower the return on investment from applying PCGML in the firstplace. In response, we propose the use of discriminative models (which capturethe validity of a design rather the distribution of the content) trained onpositive and negative examples. Through a modest modification ofWaveFunctionCollapse, a commercially-adopted PCG approach that we characterizeas using elementary machine learning, we demonstrate a new mode of control forlearning-based generators. We demonstrate how an artist might craft a focusedset of additional positive and negative examples by critique of the generator'sprevious outputs. This interaction mode bridges PCGML with mixed-initiativedesign assistance tools by working with a machine to define a space of validdesigns rather than just one new design.