Evaluation Report: iter3_mae_384d8

ITERATION 3: Switch to MAE model (no autoregressive compounding) ================================================================ Changes from Iteration 2: - Model type: GPT -> MAE (Masked Autoencoder) - Architecture: encoder-decoder with spatial masking - embed_dim=384, depth=8 (encoder), decoder_embed_dim=192, decoder_depth=4 - Mask strategy: mask last frame region, reconstruct from context - NO autoregressive generation — all patches predicted in parallel Rationale: The GPT model's biggest limitation is autoregressive error compounding. The per-position loss shows 10-50x higher error at the end of the raster scan (motor strip patches). The TF/FR gap remained at 0.006 even with 3x more capacity. MAE predicts all masked patches simultaneously, eliminating: 1. Error compounding (each patch is independent of other predictions) 2. TF/FR gap (no autoregressive generation) 3. Position-dependent quality degradation Expected improvements: - Motor strip accuracy (no longer at bottom of raster scan) - More uniform spatial error distribution - Potentially better SSIM (no blur accumulation) Potential downsides: - MAE doesn't capture sequential dependencies between patches - May produce less coherent local structure - Different failure modes (may produce "average" predictions) Comparison targets from iter2 (GPT 384d12): val_mse=0.0102, SSIM=0.756, PSNR=20.4, action_disc=0.031 motor_dir_acc=83.1%, motor_pos_mae=0.612, motor_consistency=1.01