Anthropic's Jacobian Lens enables open-source router to detect local LLM hallucinations

Anthropic recently published their Global Workspace / Jacobian Lens paper, offering a way to peek inside a model's reasoning. A developer (Kyle) quickly adapted it to open models — Gemma 4 variants (E4B, 12B, 12B abliterated, 26B MoE) and Qwen 3.6 27B — by fitting lenses that analyze how candidate answers evolve across layers. The key insight: when the model knows the answer, the workspace looks calm (one candidate dominates early, layers agree); when it is about to confidently BS, competing candidates persist deeper into the network. Testing on 500 TriviaQA questions per model showed clean workspaces were 77% correct vs 42% for noisy ones.

To make this practical, Kyle trained a simple logistic regression router on workspace trajectory features (entropy slope, late-band entropy, layer agreement, etc.). The router consistently beat raw output logprobs on all Gemma models (combined AUC up to .843 on the 12B), though it didn't help the already well-calibrated Qwen. The trained routers are released on HuggingFace, and the E4B router transfers zero-shot to other Gemmas at ~0.74-0.78 AUC. A striking side finding: abliteration (removing refusal training) made the model fabricate fake entities on 49/50 prompts vs 17/50 for the base model, highlighting a dangerous failure mode for local models.