Update Caveat reference and add corrected test_v2 scores

Add pointer to Devseis AI Act Classifier v5

Link to the live Caveat demo

Part 3: retrain on legally-grounded data, document the more-data-is-not-better regression and the macro-F1 fix

Add ONNX export (fp32 + int8 quantized) for in-browser inference via transformers.js

Add Part 2: reasoning-based approach, with self-grading caveat clearly stated

Add model card and training/eval data for transparency

Upload trained model weights (negative result experiment)

initial commit