🧠 NLA Brain-in-a-Jar v2

Type a prompt. Phi-4 14B generates a response, then an activation verbalizer describes what two key layers were computing — with confidence scoring and honest hedging when uncertain.

⚠️ You must be logged into HuggingFace (free account works!) for GPU access. Click 'Sign In' top-right if you get an error.

GRPO AV (AR-native trained), compass-reranked. Part of the NLA-at-Home project. For full layer-by-layer view, run locally: brain_in_jar_phi4.py

First call may be slow (~2 min for model loading). Subsequent calls ~30-60s.