This paper does provides a method for fitting a recurrent network to human behavioural data in an interpretable way. This is achieved via kind of auto-encoding process, where behaviors are encoded into a latent space, which are then decoded into the weights of a second RNN, which is trained to predict subject-specific data. The three expert reviewers all recommended this work for acceptance, although they pointed out some important limitations. R1 in particular was concerned about the lack of ablations; it is not immediately clear to the reader how the quantify the importance of various components of the proposed approach (such as hypernets). R2 and R1 were also concerned about the lack of comparison with any established method, and R2 was keen to see a comparison with the more immediately transparent 'cognitive model' approaches that the authors hint at in the abstract of the paper. Despite these limitations, the overall consensus was that the work is of high quality and makes important contributions to how RNNs can be applied to understand human behaviour and decision making. As such, I think it warrants acceptance to the conference.