Anthropic maps 1,200 J-space tokens in Claude 4 with no verified causal effect on output logits

Anthropic applied sparse probing to Claude 4's 70-layer transformer and extracted a subspace labeled J-space containing tokens such as 'panic' and 'protein' that activate during chain-of-thought traces. The method recovered these features from activation patches on 8,400 synthetic tasks but reported only correlational metrics; no ablation or steering vectors were tested against downstream loss. Output distributions remained unchanged when J-space dimensions were zeroed, indicating the tokens function as passive monitors rather than controllers.

The work extends Anthropic's prior dictionary-learning papers from 2023-2025 yet repeats the same limitation: monosemanticity is measured by reconstruction error on held-out activations, not by behavioral change. Comparable results appear in DeepMind's 2025 circuit papers on Gemma, where feature ablation produced measurable logit shifts only after explicit gradient-based optimization. Mainstream coverage framed the tokens as 'internal thoughts,' a description unsupported by the absence of any causal graph linking J-space to generated text.

Operationally the finding narrows the search space for future editing techniques but supplies no new control surface. Production safety pipelines at scale continue to rely on output filtering and refusal training, neither of which benefits directly from passive feature catalogs. Subsequent experiments must demonstrate at least a 10% reduction in targeted failure modes under J-space steering before the subspace can be treated as an actionable interface.