Upload app.py with huggingface_hub

Slim down for free-tier ZeroGPU (120s limit): 2 layers, best-of-2, shorter gen

Add login requirement notice + friendly GPU error message

fix: hook-based injection (matches training), greedy+sampled pool

v2: compass reranking + confidence-gated policy + universal-av-v2

Revert to v1 — post-processing makes it worse, need training fix

v3: comparative distill — extract concrete tokens, not summaries

Add single-sentence distillation post-processing

Better README with what-to-look-for guide

Forward-hook injection (post-hook on blocks[0])

Lazy model loading for ZeroGPU

Upload README.md with huggingface_hub

Upload requirements.txt with huggingface_hub

Upload app.py with huggingface_hub

Upload folder using huggingface_hub

initial commit