WAMProbe¶

WAMProbe is a counterfactual evaluation toolkit for World Action Models. It asks whether different candidate actions produce different predicted futures, whether those differences agree with the real dynamics, and whether the predictions improve action selection.

The same initial state is restored before every branch:

shared context
├── no-op       → predicted future / true future
├── move left   → predicted future / true future
├── move right  → predicted future / true future
└── expert act  → predicted future / true future

WAMProbe reports action dependence, direction correctness, physical accuracy, candidate ranking, regret, closed-loop return, uncertainty, and compute cost as separate profiles. It intentionally does not publish a single composite score.

Start here¶

• Follow the 15-minute quick start for dependency-free CPU experiments.
• Connect a model with the tested custom adapter starter guide.
• Read the scope and capability RFC before adding an adapter.
• Use the core metric cards to select capability-compatible metrics.
• Review the reproducibility guide before comparing results.
• Inspect the StarWAM model card for the first real-model integration.
• Check Next steps for milestone dependencies and contribution-sized tasks.

Install the dependency-free core with pip install wamprobe. Committed example reports and audited release artifacts remain available on PyPI and in the GitHub repository.