Paper status: preprint; arXiv submission pending.
World-model systems can spend more compute by invoking a larger predictor, but a larger predictor is not necessarily the better decision-maker for every state and task. This repository studies a simple question:
After running a Medium world model, can its predicted consequences tell us when paying for a separately trained Full model will improve the physical decision enough to justify the extra latency?
The main contribution is a candidate-complete paired exact-reset evaluation protocol. Every candidate action is executed from the same simulator state, so Medium and Full are compared on identical physical outcomes rather than different samples.
On a prospectively sealed second PushT bank of 1,600 fresh states, the prediction-interface router beats a deliberately favoured control that receives the task, current DINO features, and all five candidate actions at matched input dimension.
| Comparison | Difference in priced decision cost | 95% state-clustered interval |
|---|---|---|
| Prediction interface − current-DINO + actions | −0.002549 | [−0.002867, −0.002238] |
Negative values favour the prediction interface. All three fixed checkpoint-pair effects are negative. The control is slightly faster, but the prediction interface reduces physical regret enough to lower the priced objective by 7.9%.
paper/: the current preprint PDF and LaTeX source;src/wmroute/: reusable implementations of routing costs, information-interface value, and clustered inference utilities;scripts/: paper-facing table, figure, replay, and confirmation checks;artifacts/: compact released arrays, per-state tables, seals, manifests, and paper-facing statistics;tests/: mathematical identities, bootstrap semantics, frozen-source checks, and paper-value regression tests;docs/: scientific scope, reproducibility notes, limitations, and a static project page.
The repository supports paper-facing recomputation without a GPU or simulator. Full physical outcome regeneration still requires external simulators, encoders, checkpoints, and the original environment-specific assets.
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
python -m pip install --upgrade pip
python -m pip install -r requirements/verification.txt
python -m pip install -e .
bash scripts/run_all_checks.shExpected terminal markers include:
WMROUTE_PUBLIC_RELEASE_VERIFY_PASS
PROSPECTIVE_CONFIRMATION_VERIFICATION_PASS
WMROUTE_ALL_CHECKS_PASS
python scripts/build_main_tables.py
python scripts/build_main_figures.pyOutputs are written under generated/.
Supported by the released evidence:
- Medium and Full are not decision-ordered state by state;
- a Medium-derived prediction interface supports useful post-hoc routing;
- the sealed prediction interface beats one strong current-DINO/action alternative;
- the paired exact-reset physical estimand and paper-facing statistics are reproducible from the released records.
Not established:
- raw latency, energy, or compute saving;
- closed-loop control value;
- universal superiority of predicted futures over current-state representations;
- generality across genuinely different world-model families;
- a new generic adaptive-routing theory.
See docs/claims_and_limitations.md and the paper for the precise claim boundary.
The static page is under docs/index.html. After enabling GitHub Pages from the docs/ directory, it will be available at:
https://malodept.github.io/wmroute/
Citation metadata is provided in CITATION.cff. The arXiv identifier will be added after submission.
This release is public for inspection and reproducibility review, but no open-source licence is granted yet. See LICENSE.md and THIRD_PARTY_NOTICES.md. A standard software licence can be added after the rights review is complete.