A new audio deepfake detection model is deployed, built on W2V2 with Mamba, LoRA, and LayerNorm. The headline change is that equal error rate dropped from 7% to 4%.The other numbers moved with it. Precision is now 96% and recall 94%. Detection of NVIDIA Magpie TTS reached 99%, and accuracy on VoIP audio with packet loss improved from 88% to 94%.The packet-loss gain matters for real calls, where audio is rarely clean. The model holds up better on exactly the kind of degraded audio that phone and VoIP streams produce.