| # Train/Gen Path Mismatch — Root Cause | |
| ## Finding | |
| Training uses → : | |
| - Multi-level causal linear attention over the full chunk | |
| - Kuramoto **RK4** integrate | |
| - MoE over all positions | |
| Default generation used → : | |
| - Single-step attention (level-0 style path) | |
| - Kuramoto **one Euler step** (theta + 0.1 * derivative) | |
| - MoE on one token | |
| So the model was trained on one dynamics and decoded with another. | |
| ## Fixes | |
| 1. **Code:** Kuramoto now uses (same as train). | |
| 2. **Decode:** Prefer / sliding-window chunk gen for train-aligned logits. | |
| 3. **Decode surgery:** still useful for attractor lock; does not replace path alignment. | |
| ## Expected | |
| Aligning dynamics removes a major source of collapse/incoherence. Coherent language still depends on further digestion under stage2 dense CE. | |