J-Space introduces the Jacobian lens (J-lens), a mechanistic interpretability technique that identifies verbalizable concepts within language model activations. The method constructs sparse decompositions of internal representations to surface intermediate reasoning steps, such as planning a rhyme or working through a multi-step problem, and demonstrates causal effects through targeted interventions. It draws a structural parallel to global workspace theory from neuroscience while distinguishing the functional availability of information from subjective experience. The analysis illustrates how models organize and route information internally without resolving broader debates about machine sentience.
