All eight cells (time/pos x naive/fine x OU/neural), seeds 0-9,
trained with the pinned configuration of the arXiv revision
(vectorised generator, intervention_value_scale=1.0 matched
between training and eval, configs/continuous_ablation_grid.yaml).
These are the checkpoints behind the paper's Table 1; evals and
drivers are in the GitHub repo under results/grid_v5_eval and
scripts/train_grid_v5.sh.
