Evaluation Report: iter2_larger_gpt_384d12

ITERATION 2: Larger GPT model (more capacity) ============================================== Changes from Iteration 1: - embed_dim: 256 -> 384 (50% increase) - depth: 8 -> 12 (50% more transformer blocks) - num_heads: 8 -> 12 (matched to embed_dim divisibility) - Added gradient clipping (max_norm=1.0) via training script - Parameter count: ~6.9M -> ~22M (3x more capacity) Rationale: Iteration 1 showed that 7x more data improved all metrics substantially, confirming the model was data-starved. With 700 samples and a well-fitting model (train_val_gap only -0.002), there's room to increase capacity. The main issues from iter1: 1. TF/FR gap = 0.007 (exposure bias / error compounding) 2. SSIM = 0.70 (below 0.8 — predictions lack structural detail) 3. Motor consistency error = 1.17 (position/velocity disagree) 4. Dynamic MSE still 10x static MSE More capacity should help with: - Better learning of spatial transformations (reducing dynamic MSE) - Richer representations for accurate visual prediction (improving SSIM) - Better attention over context (potentially reducing error compounding) It will NOT fix: - Fundamental autoregressive error compounding (architectural limitation) - Motor consistency (needs explicit consistency loss)