I mapped Anthropic’s J-Space Hallucination signal across 7 datasets on Qwen3-4B to find out where it works and where it breaks

Anthropic recently published their paper on "Global Workspaces" (J-Space) inside language models, showing that looking at internal "workspace noise" (entropy) can catch hallucinations better than just looking at output logprobs. solarkyle followed up with a great open-source implementation showing it could route confident errors on TriviaQA. I wanted to see if this actually holds up as a deployable tool, so I ran a stress-test: extracting J-Space entropy from the late layers (L30-L34) of Qwen3-4B across ~