Update app.py

app.py

@@ -328,7 +328,7 @@ Crucially, the exact moment the threshold hit **31%**, performance collapsed (-2
 ### 6. Conclusion & Core Findings
 1. **Multiple-Choice Interfaces Distort Calibration:** When standard token generation heads are trapped by layout options, internal confidence drops predictably into a narrow **25% to 29% band**.
 2. **Blind Ensembles Generalize Poorly:** Standard majority voting across different inference tracks penalizes the unique correct responses hidden inside sequence likelihood strings.
-3. **The Optimal Architecture:** The most robust execution pipeline for this system is an **Unsupervised Entropy-Gate Router**. By trusting standard token choices when confidence is $\ge 29\%$, and falling back to the position-blind Perplexity engine when confidence drops below $< 29\%$, the pipeline maximizes the model's performance without degrading base performance across unseen data distributions.
+3. **The Optimal Architecture:** The most robust execution pipeline for this system is an **Unsupervised Entropy-Gate Router**. By trusting standard token choices when confidence is 29%, and falling back to the position-blind Perplexity engine when confidence drops below 29%, the pipeline maximizes the model's performance without degrading base performance across unseen data distributions.
 """)
 
     # --- Reactive Event Loop ---