When an LLM makes a decision, it typically returns a label and a confidence score. Yet neither tells us why the model decided that way, nor what kind of uncertainty lies behind the number. In this talk, I will present a way to open the model by routing its decisions through a layer of human-understandable concepts. This layer acts as an interface between perception and understanding. I will frame such models as supervised latent-variable models (SLVMs), with Concept Bottleneck Models as a concrete instantiation. Two ideas are central to this view. The first is grounding: concepts should be tied to the input spans that support them, namely the rationales. This gives explanations that say both where the model looked and why this evidence mattered. The second is uncertainty: instead of predicting a single probability for each concept, we predict an interval, or credal set, inspired by imprecise probability. This small change makes it possible to separate two different sources of uncertainty: model confusion and genuine ambiguity in the input, corresponding to epistemic and aleatoric uncertainty. This distinction is actionable. It helps decide whether to automate, collect more data, ask a human, or abstain. I close by briefly extending the imprecise view to explanation quality itself, learning to rank candidate explanations to give a richer training signal.