v4: detector fine-tuned on OpenFake with WhatsApp-style training

Weights for v4: the Community Forensics ViT-S/16 (384 px) detector, fine-tuned end to end on 13,866 OpenFake crops from 71 generators, with identical JPEG / WhatsApp-style processing for real and fake images.

Download verify_v4.pth into models/. The config (models/v4_config.json) is in the repo.

Results:

• OpenFake validation, simulated WhatsApp: AUC 0.979 (base model 0.928); held-out generators 0.921
• External set, 41 originals: 23/25 AI caught, 0/16 false alarms, AUC 0.95
• 33 real WhatsApp forwards: 12/17 AI caught, 1/16 false alarm, AUC 0.88

Threshold 0.682, chosen on validation data (5% false alarms). Training data: OpenFake (CC BY-NC 4.0). Base weights: OwensLab/commfor-model-384 (MIT).