Anthropic J-lens isolates J-space in Claude Opus 4.6 middle layers

Anthropic applied the J-lens to Claude Opus 4.6 released February 2026. The method computes Jacobian matrices between layer activations and output logits to surface words tied to tokens 3-5 steps ahead rather than immediate next-token prediction. Results were posted on the company site and integrated into a Neuronpedia demo.

Logit lens baselines from prior Anthropic work captured only immediate next-token candidates in output layers. J-space data instead surfaces thematic precursors in middle layers that often diverge from final output, including internal planning tokens absent from generated text. McGrath at Goodfire confirmed similar forward-computation patterns in independent probes.

The finding extends mechanistic interpretability beyond activation patching and sparse autoencoders by exposing explicit future-state manifolds. This directly informs alignment techniques that monitor or steer conceptual trajectories before token emission. Operational deployment could enable layer-specific interventions that reduce divergence between stated and actual model behavior.

Next steps include scaling J-lens extraction to models above 100B parameters and testing causal control via J-space ablation in production systems.