jspace-hallucination-eval

Multi-dataset evaluation of Anthropic's J-Space on Qwen3-4B. Demonstrates that internal 'workspace noise' can reliably route high-confidence confabulations, while mapping the structural limitations of the metric across adversarial, reasoning, and multiple-choice tasks.