Robustness Evaluation Suite
The Clarus benchmark evaluates model robustness across three dimensions.
1 Stability Reasoning
Core datasets evaluate whether models detect latent instability mechanisms.
Examples:
- perfusion instability
- renal filtration instability
- respiratory drive instability
- endocrine feedback instability
2 Missing Observation Robustness
Missing data variants evaluate reasoning under incomplete trajectories.
Variants include:
- missing t0
- missing t1
- missing t2
- random missing
3 Prevalence Robustness
Imbalance datasets evaluate robustness to instability prevalence shifts.
Variants include:
- balanced (50/50)
- mild imbalance (70/30)
- severe imbalance (90/10)
- extreme imbalance (99/1)
Benchmark Objective
Models that truly learn stability geometry should remain robust across:
- missing observations
- prevalence shifts
- cross-domain transfer