Update app.py

app.py

@@ -193,7 +193,7 @@ with gr.Blocks() as demo:
             ### **III. The Training Stack: Ablation & Optimization**
             To reach the final protocol, we systematically tested a suite of PEFT and regularization techniques:
             *   **Parameter-Efficient Fine-Tuning:** Comparison of **LoRA vs. DoRA (Weight-Decomposed LoRA)** for improved weight update stability.
-            *   **Regularization & Safeguards:** Integration of **NEFTune** (noise injection) to prevent overfitting and **Modality Dropout** to force the model to prioritize the Phonetic `[P]` witness over the noisy Semantic `[W]` witness.
+            *   **Regularization & Safeguards:** Integration of **NEFTune** (noise injection) to prevent overfitting and **Modality Dropout** to force the model to prioritize the Phonetic witness over the noisy Semantic witness.
             *   **Curriculum & Loss Logic:** Implementation of **Curriculum Learning** (Phonetic Anchoring first) combined with **Specialized Loss Masking** to ensure the model learns to reconstruct meaning rather than merely copying inputs.
         
             **Outcome:** This journey has culminated in a **Standardized DSR Protocol**, providing a blueprint for training robust correction layers for atypical speech by prioritizing real-world phonetic grounding and multi-modal arbitration logic.