Anthropic's J-space reveals Claude's hidden reasoning; Qwen3-8B gets same lens

Anthropic's recent research on Claude revealed a silent internal reasoning layer called J-space, where the model processes concepts in activation space without generating visible tokens. For example, Claude might answer '49' while J-space shows intermediate steps like 21 → 42 → 49, invisible to standard chain-of-thought. This distinction matters because chain-of-thought outputs readable text, while J-space captures latent reasoning—essentially what the model is 'thinking' before committing to text.

An independent developer applied the same Jacobian lens (J-lens) locally to Qwen3-8B, a 8-billion-parameter open-source model. They detected prose drift before tool calls, where the model's internal state leaned toward natural language patterns instead of structured JSON. This insight was used to build agent guards: stop or cancel undesired behavior, keep useful activation spaces, and distill recoveries into LoRA fine-tuning data. The approach offers a new technique for monitoring and controlling large language model behavior at the activation level, with potential applications in safety and tool-use reliability.