The 1.5 MB capacity cliff
A width-72, four-block transformer trained on the same passive Blocket League curriculum as the deployed model, with 10.27× fewer parameters.
Loading hallucinations…
| Metric | Value | Notes |
|---|---|---|
| 12-frame error | 4.98 px | deployed model 0.93 px |
| 64-frame error | 19.10 px | deployed model 6.53 px |
| FP32 checkpoint | 1.54 MB | 14.88 MB deployed graph |
| Training time | 4m 18s | 12,000 H100 steps |
The model learns coarse motion and the static arena, but its entities fragment under autoregressive rollout. Tenfold compression crosses the useful capacity threshold for this recipe.