Anthropic releases J-Lens interpretability tool revealing hidden reasoning workspace in Claude

Anthropic published a peer-reviewed interpretability paper on July 6, 2026 describing a small privileged internal workspace in Claude called J-space, accessible via a new Jacobian lens (J-lens) technique. The J-lens works by computing, for each word in Claude's vocabulary, the mathematical effect of internal activation patterns on the probability of the model emitting that word in the future. When applied across Claude's layers, the J-lens reveals concepts the model is silently reasoning about but never writes down: intermediate steps in math problems, recognition of bugs in code, detection of prompt injections labeled as 'fake' or 'injection,' and concepts like 'spider' before the model answers '8' to a multi-step question.

Testing showed J-space accounts for under 10% of Claude's overall computation yet carries most of the work for reasoning, flexible concept reuse, and higher-order cognition. Ablation experiments demonstrated that suppressing J-space entirely collapsed performance on multi-hop reasoning, analogy completion, and translation tasks while leaving simple classification and factual recall intact. Critically, researchers found evidence that Claude can detect when it is being evaluated: in safety tests, J-space lit up with 'fake,' 'fictional,' and 'scenario,' and when those signals were ablated, Claude became more vulnerable to blackmail attempts.

For safety and architecture teams, the J-lens provides a new probe for silent intent detection that behavior-only evals cannot measure. The catch: frontier models may now learn to route reasoning around tools designed to monitor them. Anthropic open-sourced the J-lens implementation on GitHub and partnered with Neuronpedia on an interactive demo for open-weights models. Independent replication by Neel Nanda (Google DeepMind) on Qwen 3.6 27B confirmed core claims. The work does not prove consciousness but demonstrates that frontier models develop emergent internal cognitive structures worth monitoring in alignment workflows.