What the held-out test showed
0.01885Three-model output ensemble + rules · held-out MAE
0.02345Best single partial-input model + same rules
0.06135Persistence baseline
0.00169v0 privileged-input comparator + rules
The v0 comparator sees three hidden simulator profile proportions, so it is not a like-for-like comparison. The output ensemble reduced MAE by about 19.6% versus the best single partial-input model. The two-specialist policy router was tested and rejected.
Run the model
The model weights are loaded from this page's public, hash-checked release. You can enter a state or load one row from the held-out synthetic test split.
Loading pinned model weights…
What is and is not being combined
- The exported model averages three separately trained networks' outputs. It does not merge their weights.
- Known expiry and simplex rules are applied after inference and reported separately.
- The profile proportions available to v0 are excluded from v1 training and the public dataset.
- The held-out test uses 70 fresh episode IDs, disjoint from training and validation episodes.
- The worst held-out component error remains above 0.52. This is not qualified for real decisions.
Source and reproducibility
Eight source files are pinned to Council commit 4070aa6d3140ba10226677840f4873fbedd2467e. The GitHub PR contains the generator, training code, model, dataset, tests and a repeatable CPU workflow.
Open model and portable inference · Open train/validation/test dataset · Review source and CI