A working notebook for an essay on the J-space, the small, privileged region Anthropic reported inside Claude in July 2026, and its uncanny resemblance to the “global workspace” that consciousness science has argued about for forty years.

The through-line I want to trace: a theory built to explain human consciousness, proposed as an engineering target for AI, and then found (apparently unbuilt, emergent) inside a model that was only ever trained to predict text. Three literatures that developed separately and have just collided.

Threads to follow:

• The theory. Baars invents the workspace as metaphor (1988); Dehaene and Changeux give it neurons (1998 to 2001); by 2020 it is the mainstream account of conscious access. Broadcast, limited capacity, ignition, long-range connectivity: every load-bearing term is defined here first.
• Access, not phenomenal. Block’s distinction is the hinge. Anthropic claim only access consciousness (reportable, usable, causal) and pointedly not “what it is like.” Whether that hedge holds is the whole question.
• From target to finding. VanRullen & Kanai and Goyal et al. proposed building a workspace into neural nets. Anthropic report discovering one. The inversion is the story: scarcity and long-range coordination may be a convergent solution, not a design choice.
• The instrument. J-lens descends from the logit lens and the tuned lens. The methodological weight rests on intervention (swapping, injecting, ablating concepts), not just decoding. Correlation would prove nothing; causality is the claim.
• The uncanny replications. The white-bear suppression effect, reproduced on an activation read-out. Private awareness of being tested. Detection of fabricated data. Where human psychology becomes a directly measurable experiment.

Order runs newest-first by publication date, so the 2026 primary sources sit at the top and the foundational 1987 to 2001 canon at the bottom. Read up the page for the story, down the page for the lineage.