This page runs the browser implementation of the world model against fixtures generated by the original PyTorch code, so you can confirm for yourself that the interactive demo on the main page really is the paper's model and not an approximation of it. Everything runs locally; nothing is uploaded.
Each case is a full 40-year autoregressive rollout at a real study cell. The expected values come from the PyTorch model, stored to four decimals — so an error near 5×10-5 is the rounding floor of the fixture, not model error.
| Stand age | Cell | Actions | Max abs error | 2021 height | Result |
|---|
The learned height decoder must be monotone in the action for the inverse solver to be exact. Checked on 4,001 actions from −10 m to +10 m at each stand age.
| Stand age | Min slope | Max slope | Inverse round-trip error | Result |
|---|
The digital twin's 256 cells ship packed (year-to-year deltas, split byte planes, gzip). Here the page unpacks them exactly as the twin does and rolls cells out under the twin's own tools, against the same inputs run through the paper's PyTorch code, stored to five decimals.
| Stand age | Cell | Actions | Max abs error | 2021 height | Result |
|---|
| Operation | Time | Note |
|---|