Open J-lens applied to Qwen3-8B detects silent reasoning and prose drift

The author applies an open Jacobian lens (J-lens) to the Qwen3-8B model to detect internal reasoning states, inspired by Anthropic's discovery of Claude's silent "J-space." This approach allows for monitoring hidden activations that do not appear as visible text output.

• The J-lens captures silent concepts in activations, distinguishing them from readable chain-of-thought text.
• The tool successfully detected prose drift before tool calls, identifying the model leaning toward natural language ("To, You, Do…") instead of JSON.
• The detection was integrated into agent guards to stop or cancel operations and distill recoveries into LoRA data.

This method provides a mechanism for monitoring and correcting internal model behavior that remains invisible in standard text outputs.