| # Missing Data Evaluation Protocol | |
| The Clarus benchmark includes a missing-observation evaluation suite. | |
| These datasets simulate incomplete monitoring conditions commonly observed in clinical and real-world systems. | |
| ## Missing Data Variants | |
| Four missing-data regimes are evaluated. | |
| ### Missing t1 | |
| The intermediate time point is removed. | |
| Example: | |
| t0 → missing → t2 | |
| ### Missing t2 | |
| The final observation is missing. | |
| Example: | |
| t0 → t1 → missing | |
| ### Missing t0 | |
| The initial observation is missing. | |
| Example: | |
| missing → t1 → t2 | |
| ### Random Missing | |
| One or more observations are randomly removed. | |
| ## Purpose | |
| These variants evaluate whether models can infer stability from partial trajectories. | |
| This tests robustness to incomplete observation rather than perfect monitoring. | |
| ## Evaluation | |
| The prediction task remains unchanged. | |
| Models must produce: | |
| scenario_id,prediction | |
| Evaluation uses the standard Clarus scorer. |