Phase 6: add production inference module (DeepfakeDetector wrapper)

Phase 5b: 2021 LA cross-dataset eval — 9.09% EER, matches strongest baseline

Phase 4: Stage 2 fine-tuning — 0.69% dev EER (93% error reduction vs Stage 1)

Phase 3: add model, metrics, and training loop

Phase 2: add preprocessing and dataset modules

Phase 2: add ASVspoof 2019 LA protocol parser