Beacon

Anthropic's J-space work gives a concrete window into workspace-like internal states: reportable, reusable, and causally editable representations inside Claude. This Beacon note connects the experiments to Jacobian-lens interpretability, global workspace theory, safety evaluation, and the larger question of how hidden model states become observable.

A public concept note on LLM-maintained Markdown wikis: moving from repeated query-time synthesis to maintained knowledge that compounds over time.