v1.0: research release

The pre-registered hallucination-detection campaign on Gemma-4-12B is complete.

• Gate A (cross-dataset LODO): HIT, +0.065 mean AUROC increment over logprob-only, positive 6/7 held-out datasets. Survived an independent adversarial audit and reproduction (Codex / GPT-5.5).
• Gate B (routing value at 20% budget): marginal HIT, +5.1pp catch-rate, CI [+3.2, +7.4], concentration-driven.
• Gate D (prospective zero-shot transfer to 5 never-seen datasets, classifiers frozen and hashed before generation): MISS, mean -0.016, breadth 2/5. The mapping transfers on grounded and retrieval QA (facts_grounding +0.063, nq_open +0.023) and inverts on veracity tasks (legal -0.107): fog accompanies error in retrieval and correct rejection in veracity judgment.
• Methodological warning documented in the Stage 2 report: on fixed-truth slices, internal-feature probes reach AUC ~1.0 by reading which answer the model will emit, not by detecting error.

All artifacts are public: frozen classifiers in-repo and on HF (solarkyle/jspace-lenses), 24.5k graded traces and features on HF (solarkyle/jspace-hallucination-campaign). A 90-second CPU-only reproduction is included: pip install lightgbm numpy huggingface_hub, then python campaign/reproduce_mini.py.