We do some spline carpentry to produce flexible and general representations of conditional distributions for continously supported dependent variables.We then subject the maximum likelihood estimates of these representations to various regularization strategies, including the ’stability selection’ method proposed by Meinshausen and Buhlmann (2010) and refined by Shah and Samworth (2012). The result is a semi-automated procedure that quickly produces reliable estimates of conditional distribution representations that are globally valid and are not subject to locationscale restrictions: conditional densities may be multimodal in some regions of covariate space though unimodal in others. We demonstrate the use of the techniques in two examples. Joint with Sami Stouli.