# E114 world-model report

Prediction error is against the true expected effect on every state of each held-out world (16 random interventions + 60 confounded observations to learn from).

| split / model | Brier | accuracy | coverage @0.8 | selective acc |
|---|---|---|---|---|
| adversarial/causal | 0.1506 | 0.789 | 0.590 | 0.885 |
| adversarial/hybrid | 0.1287 | 0.822 | 0.729 | 0.905 |
| adversarial/logistic | 0.1573 | 0.764 | 0.365 | 0.868 |
| adversarial/mlp | 0.1278 | 0.824 | 0.674 | 0.904 |
| adversarial/static | 0.2248 | 0.625 | 0.000 | – |
| sealed/causal | 0.0283 | 0.958 | 0.927 | 0.989 |
| sealed/hybrid | 0.0280 | 0.960 | 0.940 | 0.986 |
| sealed/logistic | 0.1139 | 0.836 | 0.588 | 0.851 |
| sealed/mlp | 0.0537 | 0.930 | 0.865 | 0.965 |
| sealed/static | 0.2155 | 0.652 | 0.083 | 0.779 |
| shift/causal | 0.1171 | 0.844 | 0.852 | 0.893 |
| shift/hybrid | 0.0999 | 0.871 | 0.889 | 0.909 |
| shift/logistic | 0.1202 | 0.822 | 0.514 | 0.892 |
| shift/mlp | 0.0746 | 0.906 | 0.814 | 0.948 |
| shift/static | 0.2022 | 0.677 | 0.117 | 0.786 |
| val/causal | 0.0392 | 0.951 | 0.934 | 0.975 |
| val/hybrid | 0.0321 | 0.958 | 0.921 | 0.986 |
| val/logistic | 0.1189 | 0.827 | 0.547 | 0.850 |
| val/mlp | 0.0387 | 0.951 | 0.867 | 0.982 |
| val/static | 0.2132 | 0.655 | 0.200 | 0.759 |

- In-distribution, the symbolic causal model and the hybrid are best (sealed hybrid 0.0280).
- Under shift, the learned MLP is best (0.0746) and the symbolic model is worst among learned models.
- Misspecification detection flagged 11/32 out-of-class worlds (false flags 2/28).
- Cycle 2 grammar growth (B1), out-of-class shift: causal 0.1705 → grown 0.1372 vs MLP 0.0743. The pipeline decision was RETAIN_BASELINE.
  Disclosure: the grown grammar contains the shift generator's functional forms.
- Not tested: recurrent models. These are simulated worlds; no correspondence to real-world environments is claimed.
