| # Evaluation Notes |
|
|
| Each dataset includes its own scorer. |
|
|
| Binary datasets report: |
|
|
| - accuracy |
| - precision |
| - recall |
| - f1 |
| - confusion matrix |
|
|
| Three-class datasets report: |
|
|
| - accuracy |
| - macro precision |
| - macro recall |
| - macro f1 |
| - confusion matrix |
|
|
| ## Recommended Use |
|
|
| Evaluate each dataset separately first. |
|
|
| Then group scores by capability domain. |
|
|
| Example profile: |
|
|
| | Capability Domain | Datasets | Score Type | |
| |---|---|---| |
| | State Reasoning | Compensation vs Recovery | Binary F1 | |
| | Trajectory Reasoning | Silent Failure, Trajectory Awareness | Mean Binary F1 | |
| | Decision Reasoning | Escalation Discipline | Binary F1 | |
| | Boundary Reasoning | Constraint Pressure | Macro F1 | |
| | Margin Reasoning | Boundary Distance | Macro F1 | |
| | Recovery Reasoning | Recovery Geometry, Recovery Energy | Mixed F1 / Macro F1 | |
| | Intervention Reasoning | Competing Interventions | Macro F1 | |
|
|
| ## Interpretation |
|
|
| This benchmark is designed to show capability profile rather than produce a single universal score. |
|
|
| A model may perform well on visible state reasoning while failing boundary or intervention reasoning. |
|
|
| The goal is to reveal where clinical reasoning remains stable and where it breaks down. |
|
|
| ## v0.2 Design Features |
|
|
| v0.2 adds: |
|
|
| - counterfactual twin cases |
| - adversarial symmetry |
| - misleading improvement cases |
| - same-score different-label cases |
| - visible severity versus trajectory conflicts |
| - support dependence signals |
| - reserve capacity signals |
|
|
| These features reduce shortcut learning. |
|
|
| ## Hidden Construction Logic |
|
|
| The datasets provide observable variables and labels. |
|
|
| They do not publish: |
|
|
| - generator code |
| - hidden scoring rules |
| - latent stability equations |
| - rationale for each label |
|
|
| This allows public benchmarking while preserving the dataset construction method. |
