The accessible entry point. Anthropic reports a "J-space" inside Claude, a small privileged region of the activations (≈10% of variance, confined to the middle layers) that behaves like Baars' global workspace: reportable, causally load-bearing in multi-step reasoning, and reusable across tasks. Start here, then read the full paper.