J-space is defined as the sparse, nonnegative combinations of Jacobian-lens directions present in an activation: the slice of a model’s internal state that maps onto concepts the model could verbalize, typically 10 to 25 active concepts accounting for 5 to 10 percent of activation variance. Anthropic’s July 2026 paper reports that this slice is reportable (the model can say what is in it), steerable on request, causally used for silent multi-step reasoning, reused across downstream tasks, and required for flexible reasoning but not for automatic fluency, the functional signatures of a global workspace.
The name comes from the Jacobian, the mathematical object used to find it. The framing deliberately parallels conscious access in humans, with the other 90-odd percent of activity playing the role of unreportable “subconscious” processing, though the paper explicitly takes no position on whether any of this involves subjective experience.
