Transfer Stability Score (TSS)
The Transfer Stability Score measures how well a model generalizes stability reasoning across different regimes.
Definition
TSS = mean(F1_transfer)
Where F1_transfer represents F1 scores from cross-regime tests.
Interpretation
High TSS indicates that a model has learned general instability reasoning.
Low TSS suggests the model relies on dataset-specific correlations.
Relationship to CSS
CSS measures average performance across datasets.
TSS measures generalization between regimes.
Both metrics should be reported when evaluating models on the Clarus Clinical Stability Benchmark.