Evaluation Protocol
Models are evaluated using three standard configurations.
Single-Dataset Evaluation
Train and evaluate on the same dataset.
Measures regime-specific performance.
Cross-Regime Transfer
Train on one regime and evaluate on another.
Measures ability to generalize stability reasoning.
Multi-Regime Training
Train on multiple datasets simultaneously.
Evaluate whether the model can learn shared stability representations.