🧠 NLA Brain-in-a-Jar v2

Type a prompt. Phi-4 14B generates a response, then a compass-reranked activation verbalizer describes what each layer was computing — with confidence scoring and honest hedging when uncertain.

New in v2: Oracle-compass reranking (+0.16 faithfulness), confidence-gated output (uncertain ≠ hallucinated), genericness penalty.

Universal AV v2 trained with contrastive GRPO (hard negatives, centered-cosine reward). Part of the NLA-at-Home project.

First call takes ~3-5 min (model loading). Subsequent calls ~1-2 min.