Upload folder using huggingface_hub

README.md

@@ -110,4 +110,10 @@ Evaluation was performed on 2,204 test samples with various segment durations. T
 | -0.02s | 84.48% | +0.02s | 88.79% |
 | -0.01s | 20.69% | +0.01s | 13.79% |
 
-> **Note**: The ±10ms shifts are at the edge of perceptual timing tolerance and are intentionally difficult to detect.
+### Analysis
+
+The performance characteristics can be directly explained by the model's physical constraints:
+
+1.  **Resolution Limit (±0.01s)**: Performance drops significantly here because the **10ms shift** is smaller than the model's temporal resolution (**~11.6ms per frame**). Sub-frame timing differences are mathematically difficult for the Convolutional Encoder to resolve.
+2.  **Optimal Zone (±0.05s to ±0.20s)**: The model achieves **100% accuracy** here. These shifts are large enough to be resolved but small enough to fit within the **~0.36s half-receptive field**. The model can simultaneously "see" the music beat and the misaligned note, enabling a direct and precise comparison.
+3.  **Field Boundary (±0.30s to ±0.50s)**: Accuracy dips slightly (to ~90%). A **0.50s shift** often pushes the note outside the receptive field of its corresponding music beat. The model can no longer compare them directly; instead, it must rely on detecting "a note without a corresponding beat" or vice-versa, which is a harder inference task (and prone to errors if the shift lands on a different valid beat).