Head-only context injection does not test contextual adaptation

Adding a new conversational feature to a frozen speech backbone and retraining only a shallow classifier is not a valid test of whether that feature helps end-to-end routing. The frozen representation may never expose the new context in a separable form, so the classifier can ignore it or overfit. Use a matched ablation first, but interpret a negative result narrowly; the proper follow-up is adapter or backbone training with train/serve-identical serialization. This limit is strongest on small, imbalanced, silver-labeled cohorts.

Applies when: New text or state context is fused into a pretrained audio-routing model. Keywords: frozen backbone context fusion classifier head LoRA routing ablation audio